1 //===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops
11 // and generates target-independent LLVM-IR.
12 // The vectorizer uses the TargetTransformInfo analysis to estimate the costs
13 // of instructions in order to estimate the profitability of vectorization.
15 // The loop vectorizer combines consecutive loop iterations into a single
16 // 'wide' iteration. After this transformation the index is incremented
17 // by the SIMD vector width, and not by one.
19 // This pass has three parts:
20 // 1. The main loop pass that drives the different parts.
21 // 2. LoopVectorizationLegality - A unit that checks for the legality
22 // of the vectorization.
23 // 3. InnerLoopVectorizer - A unit that performs the actual
24 // widening of instructions.
25 // 4. LoopVectorizationCostModel - A unit that checks for the profitability
26 // of vectorization. It decides on the optimal vector width, which
27 // can be one, if vectorization is not profitable.
29 //===----------------------------------------------------------------------===//
31 // The reduction-variable vectorization is based on the paper:
32 // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
34 // Variable uniformity checks are inspired by:
35 // Karrenberg, R. and Hack, S. Whole Function Vectorization.
37 // Other ideas/concepts are from:
38 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
40 // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of
41 // Vectorizing Compilers.
43 //===----------------------------------------------------------------------===//
45 #include "llvm/Transforms/Vectorize.h"
46 #include "llvm/ADT/DenseMap.h"
47 #include "llvm/ADT/EquivalenceClasses.h"
48 #include "llvm/ADT/Hashing.h"
49 #include "llvm/ADT/MapVector.h"
50 #include "llvm/ADT/SetVector.h"
51 #include "llvm/ADT/SmallPtrSet.h"
52 #include "llvm/ADT/SmallSet.h"
53 #include "llvm/ADT/SmallVector.h"
54 #include "llvm/ADT/Statistic.h"
55 #include "llvm/ADT/StringExtras.h"
56 #include "llvm/Analysis/AliasAnalysis.h"
57 #include "llvm/Analysis/AliasSetTracker.h"
58 #include "llvm/Analysis/BlockFrequencyInfo.h"
59 #include "llvm/Analysis/LoopInfo.h"
60 #include "llvm/Analysis/LoopIterator.h"
61 #include "llvm/Analysis/LoopPass.h"
62 #include "llvm/Analysis/ScalarEvolution.h"
63 #include "llvm/Analysis/ScalarEvolutionExpander.h"
64 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
65 #include "llvm/Analysis/TargetTransformInfo.h"
66 #include "llvm/Analysis/ValueTracking.h"
67 #include "llvm/IR/Constants.h"
68 #include "llvm/IR/DataLayout.h"
69 #include "llvm/IR/DebugInfo.h"
70 #include "llvm/IR/DerivedTypes.h"
71 #include "llvm/IR/DiagnosticInfo.h"
72 #include "llvm/IR/Dominators.h"
73 #include "llvm/IR/Function.h"
74 #include "llvm/IR/IRBuilder.h"
75 #include "llvm/IR/Instructions.h"
76 #include "llvm/IR/IntrinsicInst.h"
77 #include "llvm/IR/LLVMContext.h"
78 #include "llvm/IR/Module.h"
79 #include "llvm/IR/PatternMatch.h"
80 #include "llvm/IR/Type.h"
81 #include "llvm/IR/Value.h"
82 #include "llvm/IR/ValueHandle.h"
83 #include "llvm/IR/Verifier.h"
84 #include "llvm/Pass.h"
85 #include "llvm/Support/BranchProbability.h"
86 #include "llvm/Support/CommandLine.h"
87 #include "llvm/Support/Debug.h"
88 #include "llvm/Support/raw_ostream.h"
89 #include "llvm/Transforms/Scalar.h"
90 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
91 #include "llvm/Transforms/Utils/Local.h"
92 #include "llvm/Transforms/Utils/VectorUtils.h"
98 using namespace llvm::PatternMatch;
100 #define LV_NAME "loop-vectorize"
101 #define DEBUG_TYPE LV_NAME
103 STATISTIC(LoopsVectorized, "Number of loops vectorized");
104 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
106 static cl::opt<unsigned>
107 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
108 cl::desc("Sets the SIMD width. Zero is autoselect."));
110 static cl::opt<unsigned>
111 VectorizationUnroll("force-vector-unroll", cl::init(0), cl::Hidden,
112 cl::desc("Sets the vectorization unroll count. "
113 "Zero is autoselect."));
116 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
117 cl::desc("Enable if-conversion during vectorization."));
119 /// We don't vectorize loops with a known constant trip count below this number.
120 static cl::opt<unsigned>
121 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
123 cl::desc("Don't vectorize loops with a constant "
124 "trip count that is smaller than this "
127 /// This enables versioning on the strides of symbolically striding memory
128 /// accesses in code like the following.
129 /// for (i = 0; i < N; ++i)
130 /// A[i * Stride1] += B[i * Stride2] ...
132 /// Will be roughly translated to
133 /// if (Stride1 == 1 && Stride2 == 1) {
134 /// for (i = 0; i < N; i+=4)
138 static cl::opt<bool> EnableMemAccessVersioning(
139 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
140 cl::desc("Enable symblic stride memory access versioning"));
142 /// We don't unroll loops with a known constant trip count below this number.
143 static const unsigned TinyTripCountUnrollThreshold = 128;
145 /// When performing memory disambiguation checks at runtime do not make more
146 /// than this number of comparisons.
147 static const unsigned RuntimeMemoryCheckThreshold = 8;
149 /// Maximum simd width.
150 static const unsigned MaxVectorWidth = 64;
152 static cl::opt<unsigned> ForceTargetNumScalarRegs(
153 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
154 cl::desc("A flag that overrides the target's number of scalar registers."));
156 static cl::opt<unsigned> ForceTargetNumVectorRegs(
157 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
158 cl::desc("A flag that overrides the target's number of vector registers."));
160 /// Maximum vectorization unroll count.
161 static const unsigned MaxUnrollFactor = 16;
163 static cl::opt<unsigned> ForceTargetMaxScalarUnrollFactor(
164 "force-target-max-scalar-unroll", cl::init(0), cl::Hidden,
165 cl::desc("A flag that overrides the target's max unroll factor for scalar "
168 static cl::opt<unsigned> ForceTargetMaxVectorUnrollFactor(
169 "force-target-max-vector-unroll", cl::init(0), cl::Hidden,
170 cl::desc("A flag that overrides the target's max unroll factor for "
171 "vectorized loops."));
173 static cl::opt<unsigned> ForceTargetInstructionCost(
174 "force-target-instruction-cost", cl::init(0), cl::Hidden,
175 cl::desc("A flag that overrides the target's expected cost for "
176 "an instruction to a single constant value. Mostly "
177 "useful for getting consistent testing."));
179 static cl::opt<unsigned> SmallLoopCost(
180 "small-loop-cost", cl::init(20), cl::Hidden,
181 cl::desc("The cost of a loop that is considered 'small' by the unroller."));
183 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
184 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
185 cl::desc("Enable the use of the block frequency analysis to access PGO "
186 "heuristics minimizing code growth in cold regions and being more "
187 "aggressive in hot regions."));
189 // Runtime unroll loops for load/store throughput.
190 static cl::opt<bool> EnableLoadStoreRuntimeUnroll(
191 "enable-loadstore-runtime-unroll", cl::init(true), cl::Hidden,
192 cl::desc("Enable runtime unrolling until load/store ports are saturated"));
194 /// The number of stores in a loop that are allowed to need predication.
195 static cl::opt<unsigned> NumberOfStoresToPredicate(
196 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
197 cl::desc("Max number of stores to be predicated behind an if."));
199 static cl::opt<bool> EnableIndVarRegisterHeur(
200 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
201 cl::desc("Count the induction variable only once when unrolling"));
203 static cl::opt<bool> EnableCondStoresVectorization(
204 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
205 cl::desc("Enable if predication of stores during vectorization."));
209 // Forward declarations.
210 class LoopVectorizationLegality;
211 class LoopVectorizationCostModel;
213 /// Optimization analysis message produced during vectorization. Messages inform
214 /// the user why vectorization did not occur.
217 raw_string_ostream Out;
221 Report(Instruction *I = nullptr) : Out(Message), Instr(I) {
222 Out << "loop not vectorized: ";
225 template <typename A> Report &operator<<(const A &Value) {
230 Instruction *getInstr() { return Instr; }
232 std::string &str() { return Out.str(); }
233 operator Twine() { return Out.str(); }
236 /// InnerLoopVectorizer vectorizes loops which contain only one basic
237 /// block to a specified vectorization factor (VF).
238 /// This class performs the widening of scalars into vectors, or multiple
239 /// scalars. This class also implements the following features:
240 /// * It inserts an epilogue loop for handling loops that don't have iteration
241 /// counts that are known to be a multiple of the vectorization factor.
242 /// * It handles the code generation for reduction variables.
243 /// * Scalarization (implementation using scalars) of un-vectorizable
245 /// InnerLoopVectorizer does not perform any vectorization-legality
246 /// checks, and relies on the caller to check for the different legality
247 /// aspects. The InnerLoopVectorizer relies on the
248 /// LoopVectorizationLegality class to provide information about the induction
249 /// and reduction variables that were found to a given vectorization factor.
250 class InnerLoopVectorizer {
252 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
253 DominatorTree *DT, const DataLayout *DL,
254 const TargetLibraryInfo *TLI, unsigned VecWidth,
255 unsigned UnrollFactor)
256 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
257 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
258 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
261 // Perform the actual loop widening (vectorization).
262 void vectorize(LoopVectorizationLegality *L) {
264 // Create a new empty loop. Unlink the old loop and connect the new one.
266 // Widen each instruction in the old loop to a new one in the new loop.
267 // Use the Legality module to find the induction and reduction variables.
269 // Register the new loop and update the analysis passes.
273 virtual ~InnerLoopVectorizer() {}
276 /// A small list of PHINodes.
277 typedef SmallVector<PHINode*, 4> PhiVector;
278 /// When we unroll loops we have multiple vector values for each scalar.
279 /// This data structure holds the unrolled and vectorized values that
280 /// originated from one scalar instruction.
281 typedef SmallVector<Value*, 2> VectorParts;
283 // When we if-convert we need create edge masks. We have to cache values so
284 // that we don't end up with exponential recursion/IR.
285 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
286 VectorParts> EdgeMaskCache;
288 /// \brief Add code that checks at runtime if the accessed arrays overlap.
290 /// Returns a pair of instructions where the first element is the first
291 /// instruction generated in possibly a sequence of instructions and the
292 /// second value is the final comparator value or NULL if no check is needed.
293 std::pair<Instruction *, Instruction *> addRuntimeCheck(Instruction *Loc);
295 /// \brief Add checks for strides that where assumed to be 1.
297 /// Returns the last check instruction and the first check instruction in the
298 /// pair as (first, last).
299 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
301 /// Create an empty loop, based on the loop ranges of the old loop.
302 void createEmptyLoop();
303 /// Copy and widen the instructions from the old loop.
304 virtual void vectorizeLoop();
306 /// \brief The Loop exit block may have single value PHI nodes where the
307 /// incoming value is 'Undef'. While vectorizing we only handled real values
308 /// that were defined inside the loop. Here we fix the 'undef case'.
312 /// A helper function that computes the predicate of the block BB, assuming
313 /// that the header block of the loop is set to True. It returns the *entry*
314 /// mask for the block BB.
315 VectorParts createBlockInMask(BasicBlock *BB);
316 /// A helper function that computes the predicate of the edge between SRC
318 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
320 /// A helper function to vectorize a single BB within the innermost loop.
321 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
323 /// Vectorize a single PHINode in a block. This method handles the induction
324 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
325 /// arbitrary length vectors.
326 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
327 unsigned UF, unsigned VF, PhiVector *PV);
329 /// Insert the new loop to the loop hierarchy and pass manager
330 /// and update the analysis passes.
331 void updateAnalysis();
333 /// This instruction is un-vectorizable. Implement it as a sequence
334 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
335 /// scalarized instruction behind an if block predicated on the control
336 /// dependence of the instruction.
337 virtual void scalarizeInstruction(Instruction *Instr,
338 bool IfPredicateStore=false);
340 /// Vectorize Load and Store instructions,
341 virtual void vectorizeMemoryInstruction(Instruction *Instr);
343 /// Create a broadcast instruction. This method generates a broadcast
344 /// instruction (shuffle) for loop invariant values and for the induction
345 /// value. If this is the induction variable then we extend it to N, N+1, ...
346 /// this is needed because each iteration in the loop corresponds to a SIMD
348 virtual Value *getBroadcastInstrs(Value *V);
350 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
351 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
352 /// The sequence starts at StartIndex.
353 virtual Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
355 /// When we go over instructions in the basic block we rely on previous
356 /// values within the current basic block or on loop invariant values.
357 /// When we widen (vectorize) values we place them in the map. If the values
358 /// are not within the map, they have to be loop invariant, so we simply
359 /// broadcast them into a vector.
360 VectorParts &getVectorValue(Value *V);
362 /// Generate a shuffle sequence that will reverse the vector Vec.
363 virtual Value *reverseVector(Value *Vec);
365 /// This is a helper class that holds the vectorizer state. It maps scalar
366 /// instructions to vector instructions. When the code is 'unrolled' then
367 /// then a single scalar value is mapped to multiple vector parts. The parts
368 /// are stored in the VectorPart type.
370 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
372 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
374 /// \return True if 'Key' is saved in the Value Map.
375 bool has(Value *Key) const { return MapStorage.count(Key); }
377 /// Initializes a new entry in the map. Sets all of the vector parts to the
378 /// save value in 'Val'.
379 /// \return A reference to a vector with splat values.
380 VectorParts &splat(Value *Key, Value *Val) {
381 VectorParts &Entry = MapStorage[Key];
382 Entry.assign(UF, Val);
386 ///\return A reference to the value that is stored at 'Key'.
387 VectorParts &get(Value *Key) {
388 VectorParts &Entry = MapStorage[Key];
391 assert(Entry.size() == UF);
396 /// The unroll factor. Each entry in the map stores this number of vector
400 /// Map storage. We use std::map and not DenseMap because insertions to a
401 /// dense map invalidates its iterators.
402 std::map<Value *, VectorParts> MapStorage;
405 /// The original loop.
407 /// Scev analysis to use.
416 const DataLayout *DL;
417 /// Target Library Info.
418 const TargetLibraryInfo *TLI;
420 /// The vectorization SIMD factor to use. Each vector will have this many
425 /// The vectorization unroll factor to use. Each scalar is vectorized to this
426 /// many different vector instructions.
429 /// The builder that we use
432 // --- Vectorization state ---
434 /// The vector-loop preheader.
435 BasicBlock *LoopVectorPreHeader;
436 /// The scalar-loop preheader.
437 BasicBlock *LoopScalarPreHeader;
438 /// Middle Block between the vector and the scalar.
439 BasicBlock *LoopMiddleBlock;
440 ///The ExitBlock of the scalar loop.
441 BasicBlock *LoopExitBlock;
442 ///The vector loop body.
443 SmallVector<BasicBlock *, 4> LoopVectorBody;
444 ///The scalar loop body.
445 BasicBlock *LoopScalarBody;
446 /// A list of all bypass blocks. The first block is the entry of the loop.
447 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
449 /// The new Induction variable which was added to the new block.
451 /// The induction variable of the old basic block.
452 PHINode *OldInduction;
453 /// Holds the extended (to the widest induction type) start index.
455 /// Maps scalars to widened vectors.
457 EdgeMaskCache MaskCache;
459 LoopVectorizationLegality *Legal;
462 class InnerLoopUnroller : public InnerLoopVectorizer {
464 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
465 DominatorTree *DT, const DataLayout *DL,
466 const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
467 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
470 void scalarizeInstruction(Instruction *Instr,
471 bool IfPredicateStore = false) override;
472 void vectorizeMemoryInstruction(Instruction *Instr) override;
473 Value *getBroadcastInstrs(Value *V) override;
474 Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate) override;
475 Value *reverseVector(Value *Vec) override;
478 /// \brief Look for a meaningful debug location on the instruction or it's
480 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
485 if (I->getDebugLoc() != Empty)
488 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
489 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
490 if (OpInst->getDebugLoc() != Empty)
497 /// \brief Set the debug location in the builder using the debug location in the
499 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
500 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
501 B.SetCurrentDebugLocation(Inst->getDebugLoc());
503 B.SetCurrentDebugLocation(DebugLoc());
507 /// \return string containing a file name and a line # for the given loop.
508 static std::string getDebugLocString(const Loop *L) {
511 raw_string_ostream OS(Result);
512 const DebugLoc LoopDbgLoc = L->getStartLoc();
513 if (!LoopDbgLoc.isUnknown())
514 LoopDbgLoc.print(L->getHeader()->getContext(), OS);
516 // Just print the module name.
517 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
524 /// \brief Propagate known metadata from one instruction to another.
525 static void propagateMetadata(Instruction *To, const Instruction *From) {
526 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
527 From->getAllMetadataOtherThanDebugLoc(Metadata);
529 for (auto M : Metadata) {
530 unsigned Kind = M.first;
532 // These are safe to transfer (this is safe for TBAA, even when we
533 // if-convert, because should that metadata have had a control dependency
534 // on the condition, and thus actually aliased with some other
535 // non-speculated memory access when the condition was false, this would be
536 // caught by the runtime overlap checks).
537 if (Kind != LLVMContext::MD_tbaa &&
538 Kind != LLVMContext::MD_alias_scope &&
539 Kind != LLVMContext::MD_noalias &&
540 Kind != LLVMContext::MD_fpmath)
543 To->setMetadata(Kind, M.second);
547 /// \brief Propagate known metadata from one instruction to a vector of others.
548 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
550 if (Instruction *I = dyn_cast<Instruction>(V))
551 propagateMetadata(I, From);
554 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
555 /// to what vectorization factor.
556 /// This class does not look at the profitability of vectorization, only the
557 /// legality. This class has two main kinds of checks:
558 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
559 /// will change the order of memory accesses in a way that will change the
560 /// correctness of the program.
561 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
562 /// checks for a number of different conditions, such as the availability of a
563 /// single induction variable, that all types are supported and vectorize-able,
564 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
565 /// This class is also used by InnerLoopVectorizer for identifying
566 /// induction variable and the different reduction variables.
567 class LoopVectorizationLegality {
571 unsigned NumPredStores;
573 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, const DataLayout *DL,
574 DominatorTree *DT, TargetLibraryInfo *TLI,
575 AliasAnalysis *AA, Function *F)
576 : NumLoads(0), NumStores(0), NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
577 DT(DT), TLI(TLI), AA(AA), TheFunction(F), Induction(nullptr),
578 WidestIndTy(nullptr), HasFunNoNaNAttr(false), MaxSafeDepDistBytes(-1U) {
581 /// This enum represents the kinds of reductions that we support.
583 RK_NoReduction, ///< Not a reduction.
584 RK_IntegerAdd, ///< Sum of integers.
585 RK_IntegerMult, ///< Product of integers.
586 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
587 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
588 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
589 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
590 RK_FloatAdd, ///< Sum of floats.
591 RK_FloatMult, ///< Product of floats.
592 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
595 /// This enum represents the kinds of inductions that we support.
597 IK_NoInduction, ///< Not an induction variable.
598 IK_IntInduction, ///< Integer induction variable. Step = 1.
599 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
600 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
601 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
604 // This enum represents the kind of minmax reduction.
605 enum MinMaxReductionKind {
615 /// This struct holds information about reduction variables.
616 struct ReductionDescriptor {
617 ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr),
618 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
620 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
621 MinMaxReductionKind MK)
622 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
624 // The starting value of the reduction.
625 // It does not have to be zero!
626 TrackingVH<Value> StartValue;
627 // The instruction who's value is used outside the loop.
628 Instruction *LoopExitInstr;
629 // The kind of the reduction.
631 // If this a min/max reduction the kind of reduction.
632 MinMaxReductionKind MinMaxKind;
635 /// This POD struct holds information about a potential reduction operation.
636 struct ReductionInstDesc {
637 ReductionInstDesc(bool IsRedux, Instruction *I) :
638 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
640 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
641 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
643 // Is this instruction a reduction candidate.
645 // The last instruction in a min/max pattern (select of the select(icmp())
646 // pattern), or the current reduction instruction otherwise.
647 Instruction *PatternLastInst;
648 // If this is a min/max pattern the comparison predicate.
649 MinMaxReductionKind MinMaxKind;
652 /// This struct holds information about the memory runtime legality
653 /// check that a group of pointers do not overlap.
654 struct RuntimePointerCheck {
655 RuntimePointerCheck() : Need(false) {}
657 /// Reset the state of the pointer runtime information.
664 DependencySetId.clear();
668 /// Insert a pointer and calculate the start and end SCEVs.
669 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
670 unsigned DepSetId, unsigned ASId, ValueToValueMap &Strides);
672 /// This flag indicates if we need to add the runtime check.
674 /// Holds the pointers that we need to check.
675 SmallVector<TrackingVH<Value>, 2> Pointers;
676 /// Holds the pointer value at the beginning of the loop.
677 SmallVector<const SCEV*, 2> Starts;
678 /// Holds the pointer value at the end of the loop.
679 SmallVector<const SCEV*, 2> Ends;
680 /// Holds the information if this pointer is used for writing to memory.
681 SmallVector<bool, 2> IsWritePtr;
682 /// Holds the id of the set of pointers that could be dependent because of a
683 /// shared underlying object.
684 SmallVector<unsigned, 2> DependencySetId;
685 /// Holds the id of the disjoint alias set to which this pointer belongs.
686 SmallVector<unsigned, 2> AliasSetId;
689 /// A struct for saving information about induction variables.
690 struct InductionInfo {
691 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
692 InductionInfo() : StartValue(nullptr), IK(IK_NoInduction) {}
694 TrackingVH<Value> StartValue;
699 /// ReductionList contains the reduction descriptors for all
700 /// of the reductions that were found in the loop.
701 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
703 /// InductionList saves induction variables and maps them to the
704 /// induction descriptor.
705 typedef MapVector<PHINode*, InductionInfo> InductionList;
707 /// Returns true if it is legal to vectorize this loop.
708 /// This does not mean that it is profitable to vectorize this
709 /// loop, only that it is legal to do so.
712 /// Returns the Induction variable.
713 PHINode *getInduction() { return Induction; }
715 /// Returns the reduction variables found in the loop.
716 ReductionList *getReductionVars() { return &Reductions; }
718 /// Returns the induction variables found in the loop.
719 InductionList *getInductionVars() { return &Inductions; }
721 /// Returns the widest induction type.
722 Type *getWidestInductionType() { return WidestIndTy; }
724 /// Returns True if V is an induction variable in this loop.
725 bool isInductionVariable(const Value *V);
727 /// Return true if the block BB needs to be predicated in order for the loop
728 /// to be vectorized.
729 bool blockNeedsPredication(BasicBlock *BB);
731 /// Check if this pointer is consecutive when vectorizing. This happens
732 /// when the last index of the GEP is the induction variable, or that the
733 /// pointer itself is an induction variable.
734 /// This check allows us to vectorize A[idx] into a wide load/store.
736 /// 0 - Stride is unknown or non-consecutive.
737 /// 1 - Address is consecutive.
738 /// -1 - Address is consecutive, and decreasing.
739 int isConsecutivePtr(Value *Ptr);
741 /// Returns true if the value V is uniform within the loop.
742 bool isUniform(Value *V);
744 /// Returns true if this instruction will remain scalar after vectorization.
745 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
747 /// Returns the information that we collected about runtime memory check.
748 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
750 /// This function returns the identity element (or neutral element) for
752 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
754 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
756 bool hasStride(Value *V) { return StrideSet.count(V); }
757 bool mustCheckStrides() { return !StrideSet.empty(); }
758 SmallPtrSet<Value *, 8>::iterator strides_begin() {
759 return StrideSet.begin();
761 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
764 /// Check if a single basic block loop is vectorizable.
765 /// At this point we know that this is a loop with a constant trip count
766 /// and we only need to check individual instructions.
767 bool canVectorizeInstrs();
769 /// When we vectorize loops we may change the order in which
770 /// we read and write from memory. This method checks if it is
771 /// legal to vectorize the code, considering only memory constrains.
772 /// Returns true if the loop is vectorizable
773 bool canVectorizeMemory();
775 /// Return true if we can vectorize this loop using the IF-conversion
777 bool canVectorizeWithIfConvert();
779 /// Collect the variables that need to stay uniform after vectorization.
780 void collectLoopUniforms();
782 /// Return true if all of the instructions in the block can be speculatively
783 /// executed. \p SafePtrs is a list of addresses that are known to be legal
784 /// and we know that we can read from them without segfault.
785 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSet<Value *, 8>& SafePtrs);
787 /// Returns True, if 'Phi' is the kind of reduction variable for type
788 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
789 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
790 /// Returns a struct describing if the instruction 'I' can be a reduction
791 /// variable of type 'Kind'. If the reduction is a min/max pattern of
792 /// select(icmp()) this function advances the instruction pointer 'I' from the
793 /// compare instruction to the select instruction and stores this pointer in
794 /// 'PatternLastInst' member of the returned struct.
795 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
796 ReductionInstDesc &Desc);
797 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
798 /// pattern corresponding to a min(X, Y) or max(X, Y).
799 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
800 ReductionInstDesc &Prev);
801 /// Returns the induction kind of Phi. This function may return NoInduction
802 /// if the PHI is not an induction variable.
803 InductionKind isInductionVariable(PHINode *Phi);
805 /// \brief Collect memory access with loop invariant strides.
807 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
809 void collectStridedAcccess(Value *LoadOrStoreInst);
811 /// Report an analysis message to assist the user in diagnosing loops that are
813 void emitAnalysis(Report &Message) {
814 DebugLoc DL = TheLoop->getStartLoc();
815 if (Instruction *I = Message.getInstr())
816 DL = I->getDebugLoc();
817 emitOptimizationRemarkAnalysis(TheFunction->getContext(), DEBUG_TYPE,
818 *TheFunction, DL, Message.str());
821 /// The loop that we evaluate.
825 /// DataLayout analysis.
826 const DataLayout *DL;
829 /// Target Library Info.
830 TargetLibraryInfo *TLI;
834 Function *TheFunction;
836 // --- vectorization state --- //
838 /// Holds the integer induction variable. This is the counter of the
841 /// Holds the reduction variables.
842 ReductionList Reductions;
843 /// Holds all of the induction variables that we found in the loop.
844 /// Notice that inductions don't need to start at zero and that induction
845 /// variables can be pointers.
846 InductionList Inductions;
847 /// Holds the widest induction type encountered.
850 /// Allowed outside users. This holds the reduction
851 /// vars which can be accessed from outside the loop.
852 SmallPtrSet<Value*, 4> AllowedExit;
853 /// This set holds the variables which are known to be uniform after
855 SmallPtrSet<Instruction*, 4> Uniforms;
856 /// We need to check that all of the pointers in this list are disjoint
858 RuntimePointerCheck PtrRtCheck;
859 /// Can we assume the absence of NaNs.
860 bool HasFunNoNaNAttr;
862 unsigned MaxSafeDepDistBytes;
864 ValueToValueMap Strides;
865 SmallPtrSet<Value *, 8> StrideSet;
868 /// LoopVectorizationCostModel - estimates the expected speedups due to
870 /// In many cases vectorization is not profitable. This can happen because of
871 /// a number of reasons. In this class we mainly attempt to predict the
872 /// expected speedup/slowdowns due to the supported instruction set. We use the
873 /// TargetTransformInfo to query the different backends for the cost of
874 /// different operations.
875 class LoopVectorizationCostModel {
877 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
878 LoopVectorizationLegality *Legal,
879 const TargetTransformInfo &TTI,
880 const DataLayout *DL, const TargetLibraryInfo *TLI)
881 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
883 /// Information about vectorization costs
884 struct VectorizationFactor {
885 unsigned Width; // Vector width with best cost
886 unsigned Cost; // Cost of the loop with that width
888 /// \return The most profitable vectorization factor and the cost of that VF.
889 /// This method checks every power of two up to VF. If UserVF is not ZERO
890 /// then this vectorization factor will be selected if vectorization is
892 VectorizationFactor selectVectorizationFactor(bool OptForSize,
894 bool ForceVectorization);
896 /// \return The size (in bits) of the widest type in the code that
897 /// needs to be vectorized. We ignore values that remain scalar such as
898 /// 64 bit loop indices.
899 unsigned getWidestType();
901 /// \return The most profitable unroll factor.
902 /// If UserUF is non-zero then this method finds the best unroll-factor
903 /// based on register pressure and other parameters.
904 /// VF and LoopCost are the selected vectorization factor and the cost of the
906 unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
909 /// \brief A struct that represents some properties of the register usage
911 struct RegisterUsage {
912 /// Holds the number of loop invariant values that are used in the loop.
913 unsigned LoopInvariantRegs;
914 /// Holds the maximum number of concurrent live intervals in the loop.
915 unsigned MaxLocalUsers;
916 /// Holds the number of instructions in the loop.
917 unsigned NumInstructions;
920 /// \return information about the register usage of the loop.
921 RegisterUsage calculateRegisterUsage();
924 /// Returns the expected execution cost. The unit of the cost does
925 /// not matter because we use the 'cost' units to compare different
926 /// vector widths. The cost that is returned is *not* normalized by
927 /// the factor width.
928 unsigned expectedCost(unsigned VF);
930 /// Returns the execution time cost of an instruction for a given vector
931 /// width. Vector width of one means scalar.
932 unsigned getInstructionCost(Instruction *I, unsigned VF);
934 /// A helper function for converting Scalar types to vector types.
935 /// If the incoming type is void, we return void. If the VF is 1, we return
937 static Type* ToVectorTy(Type *Scalar, unsigned VF);
939 /// Returns whether the instruction is a load or store and will be a emitted
940 /// as a vector operation.
941 bool isConsecutiveLoadOrStore(Instruction *I);
943 /// The loop that we evaluate.
947 /// Loop Info analysis.
949 /// Vectorization legality.
950 LoopVectorizationLegality *Legal;
951 /// Vector target information.
952 const TargetTransformInfo &TTI;
953 /// Target data layout information.
954 const DataLayout *DL;
955 /// Target Library Info.
956 const TargetLibraryInfo *TLI;
959 /// Utility class for getting and setting loop vectorizer hints in the form
960 /// of loop metadata.
961 class LoopVectorizeHints {
964 FK_Undefined = -1, ///< Not selected.
965 FK_Disabled = 0, ///< Forcing disabled.
966 FK_Enabled = 1, ///< Forcing enabled.
969 LoopVectorizeHints(const Loop *L, bool DisableUnrolling)
970 : Width(VectorizationFactor),
971 Unroll(DisableUnrolling),
973 LoopID(L->getLoopID()) {
975 // force-vector-unroll overrides DisableUnrolling.
976 if (VectorizationUnroll.getNumOccurrences() > 0)
977 Unroll = VectorizationUnroll;
979 DEBUG(if (DisableUnrolling && Unroll == 1) dbgs()
980 << "LV: Unrolling disabled by the pass manager\n");
983 /// Return the loop metadata prefix.
984 static StringRef Prefix() { return "llvm.loop."; }
986 MDNode *createHint(LLVMContext &Context, StringRef Name, unsigned V) const {
987 SmallVector<Value*, 2> Vals;
988 Vals.push_back(MDString::get(Context, Name));
989 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
990 return MDNode::get(Context, Vals);
993 /// Mark the loop L as already vectorized by setting the width to 1.
994 void setAlreadyVectorized(Loop *L) {
995 LLVMContext &Context = L->getHeader()->getContext();
999 // Create a new loop id with one more operand for the already_vectorized
1000 // hint. If the loop already has a loop id then copy the existing operands.
1001 SmallVector<Value*, 4> Vals(1);
1003 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i)
1004 Vals.push_back(LoopID->getOperand(i));
1007 createHint(Context, Twine(Prefix(), "vectorize.width").str(), Width));
1009 createHint(Context, Twine(Prefix(), "interleave.count").str(), 1));
1011 MDNode *NewLoopID = MDNode::get(Context, Vals);
1012 // Set operand 0 to refer to the loop id itself.
1013 NewLoopID->replaceOperandWith(0, NewLoopID);
1015 L->setLoopID(NewLoopID);
1017 LoopID->replaceAllUsesWith(NewLoopID);
1022 std::string emitRemark() const {
1024 if (Force == LoopVectorizeHints::FK_Disabled)
1025 R << "vectorization is explicitly disabled";
1027 R << "use -Rpass-analysis=loop-vectorize for more info";
1028 if (Force == LoopVectorizeHints::FK_Enabled) {
1029 R << " (Force=true";
1031 R << ", Vector Width=" << Width;
1033 R << ", Interleave Count=" << Unroll;
1041 unsigned getWidth() const { return Width; }
1042 unsigned getUnroll() const { return Unroll; }
1043 enum ForceKind getForce() const { return Force; }
1044 MDNode *getLoopID() const { return LoopID; }
1047 /// Find hints specified in the loop metadata.
1048 void getHints(const Loop *L) {
1052 // First operand should refer to the loop id itself.
1053 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1054 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1056 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1057 const MDString *S = nullptr;
1058 SmallVector<Value*, 4> Args;
1060 // The expected hint is either a MDString or a MDNode with the first
1061 // operand a MDString.
1062 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1063 if (!MD || MD->getNumOperands() == 0)
1065 S = dyn_cast<MDString>(MD->getOperand(0));
1066 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1067 Args.push_back(MD->getOperand(i));
1069 S = dyn_cast<MDString>(LoopID->getOperand(i));
1070 assert(Args.size() == 0 && "too many arguments for MDString");
1076 // Check if the hint starts with the loop metadata prefix.
1077 StringRef Hint = S->getString();
1078 if (!Hint.startswith(Prefix()))
1080 // Remove the prefix.
1081 Hint = Hint.substr(Prefix().size(), StringRef::npos);
1083 if (Args.size() == 1)
1084 getHint(Hint, Args[0]);
1088 // Check string hint with one operand.
1089 void getHint(StringRef Hint, Value *Arg) {
1090 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
1092 unsigned Val = C->getZExtValue();
1094 if (Hint == "vectorize.width") {
1095 if (isPowerOf2_32(Val) && Val <= MaxVectorWidth)
1098 DEBUG(dbgs() << "LV: ignoring invalid width hint metadata\n");
1099 } else if (Hint == "vectorize.enable") {
1100 if (C->getBitWidth() == 1)
1101 Force = Val == 1 ? LoopVectorizeHints::FK_Enabled
1102 : LoopVectorizeHints::FK_Disabled;
1104 DEBUG(dbgs() << "LV: ignoring invalid enable hint metadata\n");
1105 } else if (Hint == "interleave.count") {
1106 if (isPowerOf2_32(Val) && Val <= MaxUnrollFactor)
1109 DEBUG(dbgs() << "LV: ignoring invalid unroll hint metadata\n");
1111 DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint << '\n');
1115 /// Vectorization width.
1117 /// Vectorization unroll factor.
1119 /// Vectorization forced
1120 enum ForceKind Force;
1125 static void emitMissedWarning(Function *F, Loop *L,
1126 const LoopVectorizeHints &LH) {
1127 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1128 L->getStartLoc(), LH.emitRemark());
1130 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1131 if (LH.getWidth() != 1)
1132 emitLoopVectorizeWarning(
1133 F->getContext(), *F, L->getStartLoc(),
1134 "failed explicitly specified loop vectorization");
1135 else if (LH.getUnroll() != 1)
1136 emitLoopInterleaveWarning(
1137 F->getContext(), *F, L->getStartLoc(),
1138 "failed explicitly specified loop interleaving");
1142 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1144 return V.push_back(&L);
1146 for (Loop *InnerL : L)
1147 addInnerLoop(*InnerL, V);
1150 /// The LoopVectorize Pass.
1151 struct LoopVectorize : public FunctionPass {
1152 /// Pass identification, replacement for typeid
1155 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1157 DisableUnrolling(NoUnrolling),
1158 AlwaysVectorize(AlwaysVectorize) {
1159 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1162 ScalarEvolution *SE;
1163 const DataLayout *DL;
1165 TargetTransformInfo *TTI;
1167 BlockFrequencyInfo *BFI;
1168 TargetLibraryInfo *TLI;
1170 bool DisableUnrolling;
1171 bool AlwaysVectorize;
1173 BlockFrequency ColdEntryFreq;
1175 bool runOnFunction(Function &F) override {
1176 SE = &getAnalysis<ScalarEvolution>();
1177 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1178 DL = DLP ? &DLP->getDataLayout() : nullptr;
1179 LI = &getAnalysis<LoopInfo>();
1180 TTI = &getAnalysis<TargetTransformInfo>();
1181 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1182 BFI = &getAnalysis<BlockFrequencyInfo>();
1183 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
1184 AA = &getAnalysis<AliasAnalysis>();
1186 // Compute some weights outside of the loop over the loops. Compute this
1187 // using a BranchProbability to re-use its scaling math.
1188 const BranchProbability ColdProb(1, 5); // 20%
1189 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1191 // If the target claims to have no vector registers don't attempt
1193 if (!TTI->getNumberOfRegisters(true))
1197 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1198 << ": Missing data layout\n");
1202 // Build up a worklist of inner-loops to vectorize. This is necessary as
1203 // the act of vectorizing or partially unrolling a loop creates new loops
1204 // and can invalidate iterators across the loops.
1205 SmallVector<Loop *, 8> Worklist;
1208 addInnerLoop(*L, Worklist);
1210 LoopsAnalyzed += Worklist.size();
1212 // Now walk the identified inner loops.
1213 bool Changed = false;
1214 while (!Worklist.empty())
1215 Changed |= processLoop(Worklist.pop_back_val());
1217 // Process each loop nest in the function.
1221 bool processLoop(Loop *L) {
1222 assert(L->empty() && "Only process inner loops.");
1225 const std::string DebugLocStr = getDebugLocString(L);
1228 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1229 << L->getHeader()->getParent()->getName() << "\" from "
1230 << DebugLocStr << "\n");
1232 LoopVectorizeHints Hints(L, DisableUnrolling);
1234 DEBUG(dbgs() << "LV: Loop hints:"
1236 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1238 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1240 : "?")) << " width=" << Hints.getWidth()
1241 << " unroll=" << Hints.getUnroll() << "\n");
1243 // Function containing loop
1244 Function *F = L->getHeader()->getParent();
1246 // Looking at the diagnostic output is the only way to determine if a loop
1247 // was vectorized (other than looking at the IR or machine code), so it
1248 // is important to generate an optimization remark for each loop. Most of
1249 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1250 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1251 // less verbose reporting vectorized loops and unvectorized loops that may
1252 // benefit from vectorization, respectively.
1254 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1255 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1256 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1257 L->getStartLoc(), Hints.emitRemark());
1261 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1262 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1263 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1264 L->getStartLoc(), Hints.emitRemark());
1268 if (Hints.getWidth() == 1 && Hints.getUnroll() == 1) {
1269 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1270 emitOptimizationRemarkAnalysis(
1271 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1272 "loop not vectorized: vector width and interleave count are "
1273 "explicitly set to 1");
1277 // Check the loop for a trip count threshold:
1278 // do not vectorize loops with a tiny trip count.
1279 BasicBlock *Latch = L->getLoopLatch();
1280 const unsigned TC = SE->getSmallConstantTripCount(L, Latch);
1281 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1282 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1283 << "This loop is not worth vectorizing.");
1284 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1285 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1287 DEBUG(dbgs() << "\n");
1288 emitOptimizationRemarkAnalysis(
1289 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1290 "vectorization is not beneficial and is not explicitly forced");
1295 // Check if it is legal to vectorize the loop.
1296 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F);
1297 if (!LVL.canVectorize()) {
1298 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1299 emitMissedWarning(F, L, Hints);
1303 // Use the cost model.
1304 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
1306 // Check the function attributes to find out if this function should be
1307 // optimized for size.
1308 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1309 F->hasFnAttribute(Attribute::OptimizeForSize);
1311 // Compute the weighted frequency of this loop being executed and see if it
1312 // is less than 20% of the function entry baseline frequency. Note that we
1313 // always have a canonical loop here because we think we *can* vectoriez.
1314 // FIXME: This is hidden behind a flag due to pervasive problems with
1315 // exactly what block frequency models.
1316 if (LoopVectorizeWithBlockFrequency) {
1317 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1318 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1319 LoopEntryFreq < ColdEntryFreq)
1323 // Check the function attributes to see if implicit floats are allowed.a
1324 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1325 // an integer loop and the vector instructions selected are purely integer
1326 // vector instructions?
1327 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1328 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1329 "attribute is used.\n");
1330 emitOptimizationRemarkAnalysis(
1331 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1332 "loop not vectorized due to NoImplicitFloat attribute");
1333 emitMissedWarning(F, L, Hints);
1337 // Select the optimal vectorization factor.
1338 const LoopVectorizationCostModel::VectorizationFactor VF =
1339 CM.selectVectorizationFactor(OptForSize, Hints.getWidth(),
1341 LoopVectorizeHints::FK_Enabled);
1343 // Select the unroll factor.
1345 CM.selectUnrollFactor(OptForSize, Hints.getUnroll(), VF.Width, VF.Cost);
1347 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1348 << DebugLocStr << '\n');
1349 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1351 if (VF.Width == 1) {
1352 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1355 emitOptimizationRemarkAnalysis(
1356 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1357 "not beneficial to vectorize and user disabled interleaving");
1360 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1362 // Report the unrolling decision.
1363 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1364 Twine("unrolled with interleaving factor " +
1366 " (vectorization not beneficial)"));
1368 // We decided not to vectorize, but we may want to unroll.
1370 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1371 Unroller.vectorize(&LVL);
1373 // If we decided that it is *legal* to vectorize the loop then do it.
1374 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1378 // Report the vectorization decision.
1379 emitOptimizationRemark(
1380 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1381 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1382 ", unrolling interleave factor: " + Twine(UF) + ")");
1385 // Mark the loop as already vectorized to avoid vectorizing again.
1386 Hints.setAlreadyVectorized(L);
1388 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1392 void getAnalysisUsage(AnalysisUsage &AU) const override {
1393 AU.addRequiredID(LoopSimplifyID);
1394 AU.addRequiredID(LCSSAID);
1395 AU.addRequired<BlockFrequencyInfo>();
1396 AU.addRequired<DominatorTreeWrapperPass>();
1397 AU.addRequired<LoopInfo>();
1398 AU.addRequired<ScalarEvolution>();
1399 AU.addRequired<TargetTransformInfo>();
1400 AU.addRequired<AliasAnalysis>();
1401 AU.addPreserved<LoopInfo>();
1402 AU.addPreserved<DominatorTreeWrapperPass>();
1403 AU.addPreserved<AliasAnalysis>();
1408 } // end anonymous namespace
1410 //===----------------------------------------------------------------------===//
1411 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1412 // LoopVectorizationCostModel.
1413 //===----------------------------------------------------------------------===//
1415 static Value *stripIntegerCast(Value *V) {
1416 if (CastInst *CI = dyn_cast<CastInst>(V))
1417 if (CI->getOperand(0)->getType()->isIntegerTy())
1418 return CI->getOperand(0);
1422 ///\brief Replaces the symbolic stride in a pointer SCEV expression by one.
1424 /// If \p OrigPtr is not null, use it to look up the stride value instead of
1426 static const SCEV *replaceSymbolicStrideSCEV(ScalarEvolution *SE,
1427 ValueToValueMap &PtrToStride,
1428 Value *Ptr, Value *OrigPtr = nullptr) {
1430 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
1432 // If there is an entry in the map return the SCEV of the pointer with the
1433 // symbolic stride replaced by one.
1434 ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
1435 if (SI != PtrToStride.end()) {
1436 Value *StrideVal = SI->second;
1439 StrideVal = stripIntegerCast(StrideVal);
1441 // Replace symbolic stride by one.
1442 Value *One = ConstantInt::get(StrideVal->getType(), 1);
1443 ValueToValueMap RewriteMap;
1444 RewriteMap[StrideVal] = One;
1447 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
1448 DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
1453 // Otherwise, just return the SCEV of the original pointer.
1454 return SE->getSCEV(Ptr);
1457 void LoopVectorizationLegality::RuntimePointerCheck::insert(
1458 ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
1459 unsigned ASId, ValueToValueMap &Strides) {
1460 // Get the stride replaced scev.
1461 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
1462 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1463 assert(AR && "Invalid addrec expression");
1464 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1465 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1466 Pointers.push_back(Ptr);
1467 Starts.push_back(AR->getStart());
1468 Ends.push_back(ScEnd);
1469 IsWritePtr.push_back(WritePtr);
1470 DependencySetId.push_back(DepSetId);
1471 AliasSetId.push_back(ASId);
1474 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1475 // We need to place the broadcast of invariant variables outside the loop.
1476 Instruction *Instr = dyn_cast<Instruction>(V);
1478 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1479 Instr->getParent()) != LoopVectorBody.end());
1480 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1482 // Place the code for broadcasting invariant variables in the new preheader.
1483 IRBuilder<>::InsertPointGuard Guard(Builder);
1485 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1487 // Broadcast the scalar into all locations in the vector.
1488 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1493 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1495 assert(Val->getType()->isVectorTy() && "Must be a vector");
1496 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1497 "Elem must be an integer");
1498 // Create the types.
1499 Type *ITy = Val->getType()->getScalarType();
1500 VectorType *Ty = cast<VectorType>(Val->getType());
1501 int VLen = Ty->getNumElements();
1502 SmallVector<Constant*, 8> Indices;
1504 // Create a vector of consecutive numbers from zero to VF.
1505 for (int i = 0; i < VLen; ++i) {
1506 int64_t Idx = Negate ? (-i) : i;
1507 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1510 // Add the consecutive indices to the vector value.
1511 Constant *Cv = ConstantVector::get(Indices);
1512 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1513 return Builder.CreateAdd(Val, Cv, "induction");
1516 /// \brief Find the operand of the GEP that should be checked for consecutive
1517 /// stores. This ignores trailing indices that have no effect on the final
1519 static unsigned getGEPInductionOperand(const DataLayout *DL,
1520 const GetElementPtrInst *Gep) {
1521 unsigned LastOperand = Gep->getNumOperands() - 1;
1522 unsigned GEPAllocSize = DL->getTypeAllocSize(
1523 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1525 // Walk backwards and try to peel off zeros.
1526 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1527 // Find the type we're currently indexing into.
1528 gep_type_iterator GEPTI = gep_type_begin(Gep);
1529 std::advance(GEPTI, LastOperand - 1);
1531 // If it's a type with the same allocation size as the result of the GEP we
1532 // can peel off the zero index.
1533 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1541 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1542 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1543 // Make sure that the pointer does not point to structs.
1544 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1547 // If this value is a pointer induction variable we know it is consecutive.
1548 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1549 if (Phi && Inductions.count(Phi)) {
1550 InductionInfo II = Inductions[Phi];
1551 if (IK_PtrInduction == II.IK)
1553 else if (IK_ReversePtrInduction == II.IK)
1557 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1561 unsigned NumOperands = Gep->getNumOperands();
1562 Value *GpPtr = Gep->getPointerOperand();
1563 // If this GEP value is a consecutive pointer induction variable and all of
1564 // the indices are constant then we know it is consecutive. We can
1565 Phi = dyn_cast<PHINode>(GpPtr);
1566 if (Phi && Inductions.count(Phi)) {
1568 // Make sure that the pointer does not point to structs.
1569 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1570 if (GepPtrType->getElementType()->isAggregateType())
1573 // Make sure that all of the index operands are loop invariant.
1574 for (unsigned i = 1; i < NumOperands; ++i)
1575 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1578 InductionInfo II = Inductions[Phi];
1579 if (IK_PtrInduction == II.IK)
1581 else if (IK_ReversePtrInduction == II.IK)
1585 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1587 // Check that all of the gep indices are uniform except for our induction
1589 for (unsigned i = 0; i != NumOperands; ++i)
1590 if (i != InductionOperand &&
1591 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1594 // We can emit wide load/stores only if the last non-zero index is the
1595 // induction variable.
1596 const SCEV *Last = nullptr;
1597 if (!Strides.count(Gep))
1598 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1600 // Because of the multiplication by a stride we can have a s/zext cast.
1601 // We are going to replace this stride by 1 so the cast is safe to ignore.
1603 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1604 // %0 = trunc i64 %indvars.iv to i32
1605 // %mul = mul i32 %0, %Stride1
1606 // %idxprom = zext i32 %mul to i64 << Safe cast.
1607 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1609 Last = replaceSymbolicStrideSCEV(SE, Strides,
1610 Gep->getOperand(InductionOperand), Gep);
1611 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1613 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1617 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1618 const SCEV *Step = AR->getStepRecurrence(*SE);
1620 // The memory is consecutive because the last index is consecutive
1621 // and all other indices are loop invariant.
1624 if (Step->isAllOnesValue())
1631 bool LoopVectorizationLegality::isUniform(Value *V) {
1632 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1635 InnerLoopVectorizer::VectorParts&
1636 InnerLoopVectorizer::getVectorValue(Value *V) {
1637 assert(V != Induction && "The new induction variable should not be used.");
1638 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1640 // If we have a stride that is replaced by one, do it here.
1641 if (Legal->hasStride(V))
1642 V = ConstantInt::get(V->getType(), 1);
1644 // If we have this scalar in the map, return it.
1645 if (WidenMap.has(V))
1646 return WidenMap.get(V);
1648 // If this scalar is unknown, assume that it is a constant or that it is
1649 // loop invariant. Broadcast V and save the value for future uses.
1650 Value *B = getBroadcastInstrs(V);
1651 return WidenMap.splat(V, B);
1654 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1655 assert(Vec->getType()->isVectorTy() && "Invalid type");
1656 SmallVector<Constant*, 8> ShuffleMask;
1657 for (unsigned i = 0; i < VF; ++i)
1658 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1660 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1661 ConstantVector::get(ShuffleMask),
1665 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1666 // Attempt to issue a wide load.
1667 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1668 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1670 assert((LI || SI) && "Invalid Load/Store instruction");
1672 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1673 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1674 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1675 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1676 // An alignment of 0 means target abi alignment. We need to use the scalar's
1677 // target abi alignment in such a case.
1679 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1680 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1681 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1682 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1684 if (SI && Legal->blockNeedsPredication(SI->getParent()))
1685 return scalarizeInstruction(Instr, true);
1687 if (ScalarAllocatedSize != VectorElementSize)
1688 return scalarizeInstruction(Instr);
1690 // If the pointer is loop invariant or if it is non-consecutive,
1691 // scalarize the load.
1692 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1693 bool Reverse = ConsecutiveStride < 0;
1694 bool UniformLoad = LI && Legal->isUniform(Ptr);
1695 if (!ConsecutiveStride || UniformLoad)
1696 return scalarizeInstruction(Instr);
1698 Constant *Zero = Builder.getInt32(0);
1699 VectorParts &Entry = WidenMap.get(Instr);
1701 // Handle consecutive loads/stores.
1702 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1703 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1704 setDebugLocFromInst(Builder, Gep);
1705 Value *PtrOperand = Gep->getPointerOperand();
1706 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1707 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1709 // Create the new GEP with the new induction variable.
1710 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1711 Gep2->setOperand(0, FirstBasePtr);
1712 Gep2->setName("gep.indvar.base");
1713 Ptr = Builder.Insert(Gep2);
1715 setDebugLocFromInst(Builder, Gep);
1716 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1717 OrigLoop) && "Base ptr must be invariant");
1719 // The last index does not have to be the induction. It can be
1720 // consecutive and be a function of the index. For example A[I+1];
1721 unsigned NumOperands = Gep->getNumOperands();
1722 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1723 // Create the new GEP with the new induction variable.
1724 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1726 for (unsigned i = 0; i < NumOperands; ++i) {
1727 Value *GepOperand = Gep->getOperand(i);
1728 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1730 // Update last index or loop invariant instruction anchored in loop.
1731 if (i == InductionOperand ||
1732 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1733 assert((i == InductionOperand ||
1734 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1735 "Must be last index or loop invariant");
1737 VectorParts &GEPParts = getVectorValue(GepOperand);
1738 Value *Index = GEPParts[0];
1739 Index = Builder.CreateExtractElement(Index, Zero);
1740 Gep2->setOperand(i, Index);
1741 Gep2->setName("gep.indvar.idx");
1744 Ptr = Builder.Insert(Gep2);
1746 // Use the induction element ptr.
1747 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1748 setDebugLocFromInst(Builder, Ptr);
1749 VectorParts &PtrVal = getVectorValue(Ptr);
1750 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1755 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1756 "We do not allow storing to uniform addresses");
1757 setDebugLocFromInst(Builder, SI);
1758 // We don't want to update the value in the map as it might be used in
1759 // another expression. So don't use a reference type for "StoredVal".
1760 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1762 for (unsigned Part = 0; Part < UF; ++Part) {
1763 // Calculate the pointer for the specific unroll-part.
1764 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1767 // If we store to reverse consecutive memory locations then we need
1768 // to reverse the order of elements in the stored value.
1769 StoredVal[Part] = reverseVector(StoredVal[Part]);
1770 // If the address is consecutive but reversed, then the
1771 // wide store needs to start at the last vector element.
1772 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1773 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1776 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1777 DataTy->getPointerTo(AddressSpace));
1779 Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
1780 propagateMetadata(NewSI, SI);
1786 assert(LI && "Must have a load instruction");
1787 setDebugLocFromInst(Builder, LI);
1788 for (unsigned Part = 0; Part < UF; ++Part) {
1789 // Calculate the pointer for the specific unroll-part.
1790 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1793 // If the address is consecutive but reversed, then the
1794 // wide store needs to start at the last vector element.
1795 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1796 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1799 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1800 DataTy->getPointerTo(AddressSpace));
1801 LoadInst *NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
1802 propagateMetadata(NewLI, LI);
1803 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
1807 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1808 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1809 // Holds vector parameters or scalars, in case of uniform vals.
1810 SmallVector<VectorParts, 4> Params;
1812 setDebugLocFromInst(Builder, Instr);
1814 // Find all of the vectorized parameters.
1815 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1816 Value *SrcOp = Instr->getOperand(op);
1818 // If we are accessing the old induction variable, use the new one.
1819 if (SrcOp == OldInduction) {
1820 Params.push_back(getVectorValue(SrcOp));
1824 // Try using previously calculated values.
1825 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1827 // If the src is an instruction that appeared earlier in the basic block
1828 // then it should already be vectorized.
1829 if (SrcInst && OrigLoop->contains(SrcInst)) {
1830 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1831 // The parameter is a vector value from earlier.
1832 Params.push_back(WidenMap.get(SrcInst));
1834 // The parameter is a scalar from outside the loop. Maybe even a constant.
1835 VectorParts Scalars;
1836 Scalars.append(UF, SrcOp);
1837 Params.push_back(Scalars);
1841 assert(Params.size() == Instr->getNumOperands() &&
1842 "Invalid number of operands");
1844 // Does this instruction return a value ?
1845 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1847 Value *UndefVec = IsVoidRetTy ? nullptr :
1848 UndefValue::get(VectorType::get(Instr->getType(), VF));
1849 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1850 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1852 Instruction *InsertPt = Builder.GetInsertPoint();
1853 BasicBlock *IfBlock = Builder.GetInsertBlock();
1854 BasicBlock *CondBlock = nullptr;
1857 Loop *VectorLp = nullptr;
1858 if (IfPredicateStore) {
1859 assert(Instr->getParent()->getSinglePredecessor() &&
1860 "Only support single predecessor blocks");
1861 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1862 Instr->getParent());
1863 VectorLp = LI->getLoopFor(IfBlock);
1864 assert(VectorLp && "Must have a loop for this block");
1867 // For each vector unroll 'part':
1868 for (unsigned Part = 0; Part < UF; ++Part) {
1869 // For each scalar that we create:
1870 for (unsigned Width = 0; Width < VF; ++Width) {
1873 Value *Cmp = nullptr;
1874 if (IfPredicateStore) {
1875 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1876 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1877 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1878 LoopVectorBody.push_back(CondBlock);
1879 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
1880 // Update Builder with newly created basic block.
1881 Builder.SetInsertPoint(InsertPt);
1884 Instruction *Cloned = Instr->clone();
1886 Cloned->setName(Instr->getName() + ".cloned");
1887 // Replace the operands of the cloned instructions with extracted scalars.
1888 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1889 Value *Op = Params[op][Part];
1890 // Param is a vector. Need to extract the right lane.
1891 if (Op->getType()->isVectorTy())
1892 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1893 Cloned->setOperand(op, Op);
1896 // Place the cloned scalar in the new loop.
1897 Builder.Insert(Cloned);
1899 // If the original scalar returns a value we need to place it in a vector
1900 // so that future users will be able to use it.
1902 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1903 Builder.getInt32(Width));
1905 if (IfPredicateStore) {
1906 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1907 LoopVectorBody.push_back(NewIfBlock);
1908 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
1909 Builder.SetInsertPoint(InsertPt);
1910 Instruction *OldBr = IfBlock->getTerminator();
1911 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1912 OldBr->eraseFromParent();
1913 IfBlock = NewIfBlock;
1919 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1923 if (Instruction *I = dyn_cast<Instruction>(V))
1924 return I->getParent() == Loc->getParent() ? I : nullptr;
1928 std::pair<Instruction *, Instruction *>
1929 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
1930 Instruction *tnullptr = nullptr;
1931 if (!Legal->mustCheckStrides())
1932 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1934 IRBuilder<> ChkBuilder(Loc);
1937 Value *Check = nullptr;
1938 Instruction *FirstInst = nullptr;
1939 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
1940 SE = Legal->strides_end();
1942 Value *Ptr = stripIntegerCast(*SI);
1943 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
1945 // Store the first instruction we create.
1946 FirstInst = getFirstInst(FirstInst, C, Loc);
1948 Check = ChkBuilder.CreateOr(Check, C);
1953 // We have to do this trickery because the IRBuilder might fold the check to a
1954 // constant expression in which case there is no Instruction anchored in a
1956 LLVMContext &Ctx = Loc->getContext();
1957 Instruction *TheCheck =
1958 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
1959 ChkBuilder.Insert(TheCheck, "stride.not.one");
1960 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
1962 return std::make_pair(FirstInst, TheCheck);
1965 std::pair<Instruction *, Instruction *>
1966 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
1967 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
1968 Legal->getRuntimePointerCheck();
1970 Instruction *tnullptr = nullptr;
1971 if (!PtrRtCheck->Need)
1972 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1974 unsigned NumPointers = PtrRtCheck->Pointers.size();
1975 SmallVector<TrackingVH<Value> , 2> Starts;
1976 SmallVector<TrackingVH<Value> , 2> Ends;
1978 LLVMContext &Ctx = Loc->getContext();
1979 SCEVExpander Exp(*SE, "induction");
1980 Instruction *FirstInst = nullptr;
1982 for (unsigned i = 0; i < NumPointers; ++i) {
1983 Value *Ptr = PtrRtCheck->Pointers[i];
1984 const SCEV *Sc = SE->getSCEV(Ptr);
1986 if (SE->isLoopInvariant(Sc, OrigLoop)) {
1987 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
1989 Starts.push_back(Ptr);
1990 Ends.push_back(Ptr);
1992 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
1993 unsigned AS = Ptr->getType()->getPointerAddressSpace();
1995 // Use this type for pointer arithmetic.
1996 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
1998 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
1999 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
2000 Starts.push_back(Start);
2001 Ends.push_back(End);
2005 IRBuilder<> ChkBuilder(Loc);
2006 // Our instructions might fold to a constant.
2007 Value *MemoryRuntimeCheck = nullptr;
2008 for (unsigned i = 0; i < NumPointers; ++i) {
2009 for (unsigned j = i+1; j < NumPointers; ++j) {
2010 // No need to check if two readonly pointers intersect.
2011 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
2014 // Only need to check pointers between two different dependency sets.
2015 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
2017 // Only need to check pointers in the same alias set.
2018 if (PtrRtCheck->AliasSetId[i] != PtrRtCheck->AliasSetId[j])
2021 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
2022 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
2024 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
2025 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
2026 "Trying to bounds check pointers with different address spaces");
2028 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
2029 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
2031 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
2032 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
2033 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
2034 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
2036 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
2037 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
2038 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
2039 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
2040 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
2041 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2042 if (MemoryRuntimeCheck) {
2043 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
2045 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2047 MemoryRuntimeCheck = IsConflict;
2051 // We have to do this trickery because the IRBuilder might fold the check to a
2052 // constant expression in which case there is no Instruction anchored in a
2054 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
2055 ConstantInt::getTrue(Ctx));
2056 ChkBuilder.Insert(Check, "memcheck.conflict");
2057 FirstInst = getFirstInst(FirstInst, Check, Loc);
2058 return std::make_pair(FirstInst, Check);
2061 void InnerLoopVectorizer::createEmptyLoop() {
2063 In this function we generate a new loop. The new loop will contain
2064 the vectorized instructions while the old loop will continue to run the
2067 [ ] <-- Back-edge taken count overflow check.
2070 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2073 || [ ] <-- vector pre header.
2077 || [ ]_| <-- vector loop.
2080 | >[ ] <--- middle-block.
2083 -|- >[ ] <--- new preheader.
2087 | [ ]_| <-- old scalar loop to handle remainder.
2090 >[ ] <-- exit block.
2094 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2095 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2096 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2097 assert(BypassBlock && "Invalid loop structure");
2098 assert(ExitBlock && "Must have an exit block");
2100 // Some loops have a single integer induction variable, while other loops
2101 // don't. One example is c++ iterators that often have multiple pointer
2102 // induction variables. In the code below we also support a case where we
2103 // don't have a single induction variable.
2104 OldInduction = Legal->getInduction();
2105 Type *IdxTy = Legal->getWidestInductionType();
2107 // Find the loop boundaries.
2108 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2109 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2111 // The exit count might have the type of i64 while the phi is i32. This can
2112 // happen if we have an induction variable that is sign extended before the
2113 // compare. The only way that we get a backedge taken count is that the
2114 // induction variable was signed and as such will not overflow. In such a case
2115 // truncation is legal.
2116 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2117 IdxTy->getPrimitiveSizeInBits())
2118 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2120 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2121 // Get the total trip count from the count by adding 1.
2122 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2123 SE->getConstant(BackedgeTakeCount->getType(), 1));
2125 // Expand the trip count and place the new instructions in the preheader.
2126 // Notice that the pre-header does not change, only the loop body.
2127 SCEVExpander Exp(*SE, "induction");
2129 // We need to test whether the backedge-taken count is uint##_max. Adding one
2130 // to it will cause overflow and an incorrect loop trip count in the vector
2131 // body. In case of overflow we want to directly jump to the scalar remainder
2133 Value *BackedgeCount =
2134 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2135 BypassBlock->getTerminator());
2136 if (BackedgeCount->getType()->isPointerTy())
2137 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2138 "backedge.ptrcnt.to.int",
2139 BypassBlock->getTerminator());
2140 Instruction *CheckBCOverflow =
2141 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2142 Constant::getAllOnesValue(BackedgeCount->getType()),
2143 "backedge.overflow", BypassBlock->getTerminator());
2145 // The loop index does not have to start at Zero. Find the original start
2146 // value from the induction PHI node. If we don't have an induction variable
2147 // then we know that it starts at zero.
2148 Builder.SetInsertPoint(BypassBlock->getTerminator());
2149 Value *StartIdx = ExtendedIdx = OldInduction ?
2150 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2152 ConstantInt::get(IdxTy, 0);
2154 // We need an instruction to anchor the overflow check on. StartIdx needs to
2155 // be defined before the overflow check branch. Because the scalar preheader
2156 // is going to merge the start index and so the overflow branch block needs to
2157 // contain a definition of the start index.
2158 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2159 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2160 BypassBlock->getTerminator());
2162 // Count holds the overall loop count (N).
2163 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2164 BypassBlock->getTerminator());
2166 LoopBypassBlocks.push_back(BypassBlock);
2168 // Split the single block loop into the two loop structure described above.
2169 BasicBlock *VectorPH =
2170 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2171 BasicBlock *VecBody =
2172 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2173 BasicBlock *MiddleBlock =
2174 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2175 BasicBlock *ScalarPH =
2176 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2178 // Create and register the new vector loop.
2179 Loop* Lp = new Loop();
2180 Loop *ParentLoop = OrigLoop->getParentLoop();
2182 // Insert the new loop into the loop nest and register the new basic blocks
2183 // before calling any utilities such as SCEV that require valid LoopInfo.
2185 ParentLoop->addChildLoop(Lp);
2186 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
2187 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
2188 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
2190 LI->addTopLevelLoop(Lp);
2192 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
2194 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2196 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2198 // Generate the induction variable.
2199 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2200 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2201 // The loop step is equal to the vectorization factor (num of SIMD elements)
2202 // times the unroll factor (num of SIMD instructions).
2203 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2205 // This is the IR builder that we use to add all of the logic for bypassing
2206 // the new vector loop.
2207 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2208 setDebugLocFromInst(BypassBuilder,
2209 getDebugLocFromInstOrOperands(OldInduction));
2211 // We may need to extend the index in case there is a type mismatch.
2212 // We know that the count starts at zero and does not overflow.
2213 if (Count->getType() != IdxTy) {
2214 // The exit count can be of pointer type. Convert it to the correct
2216 if (ExitCount->getType()->isPointerTy())
2217 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2219 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2222 // Add the start index to the loop count to get the new end index.
2223 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2225 // Now we need to generate the expression for N - (N % VF), which is
2226 // the part that the vectorized body will execute.
2227 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2228 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2229 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2230 "end.idx.rnd.down");
2232 // Now, compare the new count to zero. If it is zero skip the vector loop and
2233 // jump to the scalar loop.
2235 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2237 BasicBlock *LastBypassBlock = BypassBlock;
2239 // Generate code to check that the loops trip count that we computed by adding
2240 // one to the backedge-taken count will not overflow.
2242 auto PastOverflowCheck =
2243 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2244 BasicBlock *CheckBlock =
2245 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2247 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2248 LoopBypassBlocks.push_back(CheckBlock);
2249 Instruction *OldTerm = LastBypassBlock->getTerminator();
2250 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2251 OldTerm->eraseFromParent();
2252 LastBypassBlock = CheckBlock;
2255 // Generate the code to check that the strides we assumed to be one are really
2256 // one. We want the new basic block to start at the first instruction in a
2257 // sequence of instructions that form a check.
2258 Instruction *StrideCheck;
2259 Instruction *FirstCheckInst;
2260 std::tie(FirstCheckInst, StrideCheck) =
2261 addStrideCheck(LastBypassBlock->getTerminator());
2263 // Create a new block containing the stride check.
2264 BasicBlock *CheckBlock =
2265 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2267 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2268 LoopBypassBlocks.push_back(CheckBlock);
2270 // Replace the branch into the memory check block with a conditional branch
2271 // for the "few elements case".
2272 Instruction *OldTerm = LastBypassBlock->getTerminator();
2273 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2274 OldTerm->eraseFromParent();
2277 LastBypassBlock = CheckBlock;
2280 // Generate the code that checks in runtime if arrays overlap. We put the
2281 // checks into a separate block to make the more common case of few elements
2283 Instruction *MemRuntimeCheck;
2284 std::tie(FirstCheckInst, MemRuntimeCheck) =
2285 addRuntimeCheck(LastBypassBlock->getTerminator());
2286 if (MemRuntimeCheck) {
2287 // Create a new block containing the memory check.
2288 BasicBlock *CheckBlock =
2289 LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
2291 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2292 LoopBypassBlocks.push_back(CheckBlock);
2294 // Replace the branch into the memory check block with a conditional branch
2295 // for the "few elements case".
2296 Instruction *OldTerm = LastBypassBlock->getTerminator();
2297 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2298 OldTerm->eraseFromParent();
2300 Cmp = MemRuntimeCheck;
2301 LastBypassBlock = CheckBlock;
2304 LastBypassBlock->getTerminator()->eraseFromParent();
2305 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2308 // We are going to resume the execution of the scalar loop.
2309 // Go over all of the induction variables that we found and fix the
2310 // PHIs that are left in the scalar version of the loop.
2311 // The starting values of PHI nodes depend on the counter of the last
2312 // iteration in the vectorized loop.
2313 // If we come from a bypass edge then we need to start from the original
2316 // This variable saves the new starting index for the scalar loop.
2317 PHINode *ResumeIndex = nullptr;
2318 LoopVectorizationLegality::InductionList::iterator I, E;
2319 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2320 // Set builder to point to last bypass block.
2321 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2322 for (I = List->begin(), E = List->end(); I != E; ++I) {
2323 PHINode *OrigPhi = I->first;
2324 LoopVectorizationLegality::InductionInfo II = I->second;
2326 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2327 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2328 MiddleBlock->getTerminator());
2329 // We might have extended the type of the induction variable but we need a
2330 // truncated version for the scalar loop.
2331 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2332 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2333 MiddleBlock->getTerminator()) : nullptr;
2335 // Create phi nodes to merge from the backedge-taken check block.
2336 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2337 ScalarPH->getTerminator());
2338 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2340 PHINode *BCTruncResumeVal = nullptr;
2341 if (OrigPhi == OldInduction) {
2343 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2344 ScalarPH->getTerminator());
2345 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2348 Value *EndValue = nullptr;
2350 case LoopVectorizationLegality::IK_NoInduction:
2351 llvm_unreachable("Unknown induction");
2352 case LoopVectorizationLegality::IK_IntInduction: {
2353 // Handle the integer induction counter.
2354 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2356 // We have the canonical induction variable.
2357 if (OrigPhi == OldInduction) {
2358 // Create a truncated version of the resume value for the scalar loop,
2359 // we might have promoted the type to a larger width.
2361 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2362 // The new PHI merges the original incoming value, in case of a bypass,
2363 // or the value at the end of the vectorized loop.
2364 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2365 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2366 TruncResumeVal->addIncoming(EndValue, VecBody);
2368 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2370 // We know what the end value is.
2371 EndValue = IdxEndRoundDown;
2372 // We also know which PHI node holds it.
2373 ResumeIndex = ResumeVal;
2377 // Not the canonical induction variable - add the vector loop count to the
2379 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2380 II.StartValue->getType(),
2382 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
2385 case LoopVectorizationLegality::IK_ReverseIntInduction: {
2386 // Convert the CountRoundDown variable to the PHI size.
2387 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2388 II.StartValue->getType(),
2390 // Handle reverse integer induction counter.
2391 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
2394 case LoopVectorizationLegality::IK_PtrInduction: {
2395 // For pointer induction variables, calculate the offset using
2397 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
2401 case LoopVectorizationLegality::IK_ReversePtrInduction: {
2402 // The value at the end of the loop for the reverse pointer is calculated
2403 // by creating a GEP with a negative index starting from the start value.
2404 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
2405 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
2407 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
2413 // The new PHI merges the original incoming value, in case of a bypass,
2414 // or the value at the end of the vectorized loop.
2415 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2416 if (OrigPhi == OldInduction)
2417 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2419 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2421 ResumeVal->addIncoming(EndValue, VecBody);
2423 // Fix the scalar body counter (PHI node).
2424 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2426 // The old induction's phi node in the scalar body needs the truncated
2428 if (OrigPhi == OldInduction) {
2429 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2430 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2432 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2433 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2437 // If we are generating a new induction variable then we also need to
2438 // generate the code that calculates the exit value. This value is not
2439 // simply the end of the counter because we may skip the vectorized body
2440 // in case of a runtime check.
2442 assert(!ResumeIndex && "Unexpected resume value found");
2443 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2444 MiddleBlock->getTerminator());
2445 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2446 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2447 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2450 // Make sure that we found the index where scalar loop needs to continue.
2451 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2452 "Invalid resume Index");
2454 // Add a check in the middle block to see if we have completed
2455 // all of the iterations in the first vector loop.
2456 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2457 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2458 ResumeIndex, "cmp.n",
2459 MiddleBlock->getTerminator());
2461 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2462 // Remove the old terminator.
2463 MiddleBlock->getTerminator()->eraseFromParent();
2465 // Create i+1 and fill the PHINode.
2466 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2467 Induction->addIncoming(StartIdx, VectorPH);
2468 Induction->addIncoming(NextIdx, VecBody);
2469 // Create the compare.
2470 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2471 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2473 // Now we have two terminators. Remove the old one from the block.
2474 VecBody->getTerminator()->eraseFromParent();
2476 // Get ready to start creating new instructions into the vectorized body.
2477 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2480 LoopVectorPreHeader = VectorPH;
2481 LoopScalarPreHeader = ScalarPH;
2482 LoopMiddleBlock = MiddleBlock;
2483 LoopExitBlock = ExitBlock;
2484 LoopVectorBody.push_back(VecBody);
2485 LoopScalarBody = OldBasicBlock;
2487 LoopVectorizeHints Hints(Lp, true);
2488 Hints.setAlreadyVectorized(Lp);
2491 /// This function returns the identity element (or neutral element) for
2492 /// the operation K.
2494 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2499 // Adding, Xoring, Oring zero to a number does not change it.
2500 return ConstantInt::get(Tp, 0);
2501 case RK_IntegerMult:
2502 // Multiplying a number by 1 does not change it.
2503 return ConstantInt::get(Tp, 1);
2505 // AND-ing a number with an all-1 value does not change it.
2506 return ConstantInt::get(Tp, -1, true);
2508 // Multiplying a number by 1 does not change it.
2509 return ConstantFP::get(Tp, 1.0L);
2511 // Adding zero to a number does not change it.
2512 return ConstantFP::get(Tp, 0.0L);
2514 llvm_unreachable("Unknown reduction kind");
2518 /// This function translates the reduction kind to an LLVM binary operator.
2520 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2522 case LoopVectorizationLegality::RK_IntegerAdd:
2523 return Instruction::Add;
2524 case LoopVectorizationLegality::RK_IntegerMult:
2525 return Instruction::Mul;
2526 case LoopVectorizationLegality::RK_IntegerOr:
2527 return Instruction::Or;
2528 case LoopVectorizationLegality::RK_IntegerAnd:
2529 return Instruction::And;
2530 case LoopVectorizationLegality::RK_IntegerXor:
2531 return Instruction::Xor;
2532 case LoopVectorizationLegality::RK_FloatMult:
2533 return Instruction::FMul;
2534 case LoopVectorizationLegality::RK_FloatAdd:
2535 return Instruction::FAdd;
2536 case LoopVectorizationLegality::RK_IntegerMinMax:
2537 return Instruction::ICmp;
2538 case LoopVectorizationLegality::RK_FloatMinMax:
2539 return Instruction::FCmp;
2541 llvm_unreachable("Unknown reduction operation");
2545 Value *createMinMaxOp(IRBuilder<> &Builder,
2546 LoopVectorizationLegality::MinMaxReductionKind RK,
2549 CmpInst::Predicate P = CmpInst::ICMP_NE;
2552 llvm_unreachable("Unknown min/max reduction kind");
2553 case LoopVectorizationLegality::MRK_UIntMin:
2554 P = CmpInst::ICMP_ULT;
2556 case LoopVectorizationLegality::MRK_UIntMax:
2557 P = CmpInst::ICMP_UGT;
2559 case LoopVectorizationLegality::MRK_SIntMin:
2560 P = CmpInst::ICMP_SLT;
2562 case LoopVectorizationLegality::MRK_SIntMax:
2563 P = CmpInst::ICMP_SGT;
2565 case LoopVectorizationLegality::MRK_FloatMin:
2566 P = CmpInst::FCMP_OLT;
2568 case LoopVectorizationLegality::MRK_FloatMax:
2569 P = CmpInst::FCMP_OGT;
2574 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2575 RK == LoopVectorizationLegality::MRK_FloatMax)
2576 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2578 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2580 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2585 struct CSEDenseMapInfo {
2586 static bool canHandle(Instruction *I) {
2587 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2588 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2590 static inline Instruction *getEmptyKey() {
2591 return DenseMapInfo<Instruction *>::getEmptyKey();
2593 static inline Instruction *getTombstoneKey() {
2594 return DenseMapInfo<Instruction *>::getTombstoneKey();
2596 static unsigned getHashValue(Instruction *I) {
2597 assert(canHandle(I) && "Unknown instruction!");
2598 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2599 I->value_op_end()));
2601 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2602 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2603 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2605 return LHS->isIdenticalTo(RHS);
2610 /// \brief Check whether this block is a predicated block.
2611 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2612 /// = ...; " blocks. We start with one vectorized basic block. For every
2613 /// conditional block we split this vectorized block. Therefore, every second
2614 /// block will be a predicated one.
2615 static bool isPredicatedBlock(unsigned BlockNum) {
2616 return BlockNum % 2;
2619 ///\brief Perform cse of induction variable instructions.
2620 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2621 // Perform simple cse.
2622 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2623 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2624 BasicBlock *BB = BBs[i];
2625 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2626 Instruction *In = I++;
2628 if (!CSEDenseMapInfo::canHandle(In))
2631 // Check if we can replace this instruction with any of the
2632 // visited instructions.
2633 if (Instruction *V = CSEMap.lookup(In)) {
2634 In->replaceAllUsesWith(V);
2635 In->eraseFromParent();
2638 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2639 // ...;" blocks for predicated stores. Every second block is a predicated
2641 if (isPredicatedBlock(i))
2649 /// \brief Adds a 'fast' flag to floating point operations.
2650 static Value *addFastMathFlag(Value *V) {
2651 if (isa<FPMathOperator>(V)){
2652 FastMathFlags Flags;
2653 Flags.setUnsafeAlgebra();
2654 cast<Instruction>(V)->setFastMathFlags(Flags);
2659 void InnerLoopVectorizer::vectorizeLoop() {
2660 //===------------------------------------------------===//
2662 // Notice: any optimization or new instruction that go
2663 // into the code below should be also be implemented in
2666 //===------------------------------------------------===//
2667 Constant *Zero = Builder.getInt32(0);
2669 // In order to support reduction variables we need to be able to vectorize
2670 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2671 // stages. First, we create a new vector PHI node with no incoming edges.
2672 // We use this value when we vectorize all of the instructions that use the
2673 // PHI. Next, after all of the instructions in the block are complete we
2674 // add the new incoming edges to the PHI. At this point all of the
2675 // instructions in the basic block are vectorized, so we can use them to
2676 // construct the PHI.
2677 PhiVector RdxPHIsToFix;
2679 // Scan the loop in a topological order to ensure that defs are vectorized
2681 LoopBlocksDFS DFS(OrigLoop);
2684 // Vectorize all of the blocks in the original loop.
2685 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2686 be = DFS.endRPO(); bb != be; ++bb)
2687 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2689 // At this point every instruction in the original loop is widened to
2690 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2691 // that we vectorized. The PHI nodes are currently empty because we did
2692 // not want to introduce cycles. Notice that the remaining PHI nodes
2693 // that we need to fix are reduction variables.
2695 // Create the 'reduced' values for each of the induction vars.
2696 // The reduced values are the vector values that we scalarize and combine
2697 // after the loop is finished.
2698 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2700 PHINode *RdxPhi = *it;
2701 assert(RdxPhi && "Unable to recover vectorized PHI");
2703 // Find the reduction variable descriptor.
2704 assert(Legal->getReductionVars()->count(RdxPhi) &&
2705 "Unable to find the reduction variable");
2706 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2707 (*Legal->getReductionVars())[RdxPhi];
2709 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2711 // We need to generate a reduction vector from the incoming scalar.
2712 // To do so, we need to generate the 'identity' vector and override
2713 // one of the elements with the incoming scalar reduction. We need
2714 // to do it in the vector-loop preheader.
2715 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2717 // This is the vector-clone of the value that leaves the loop.
2718 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2719 Type *VecTy = VectorExit[0]->getType();
2721 // Find the reduction identity variable. Zero for addition, or, xor,
2722 // one for multiplication, -1 for And.
2725 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2726 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2727 // MinMax reduction have the start value as their identify.
2729 VectorStart = Identity = RdxDesc.StartValue;
2731 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2736 // Handle other reduction kinds:
2738 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2739 VecTy->getScalarType());
2742 // This vector is the Identity vector where the first element is the
2743 // incoming scalar reduction.
2744 VectorStart = RdxDesc.StartValue;
2746 Identity = ConstantVector::getSplat(VF, Iden);
2748 // This vector is the Identity vector where the first element is the
2749 // incoming scalar reduction.
2750 VectorStart = Builder.CreateInsertElement(Identity,
2751 RdxDesc.StartValue, Zero);
2755 // Fix the vector-loop phi.
2756 // We created the induction variable so we know that the
2757 // preheader is the first entry.
2758 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2760 // Reductions do not have to start at zero. They can start with
2761 // any loop invariant values.
2762 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2763 BasicBlock *Latch = OrigLoop->getLoopLatch();
2764 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2765 VectorParts &Val = getVectorValue(LoopVal);
2766 for (unsigned part = 0; part < UF; ++part) {
2767 // Make sure to add the reduction stat value only to the
2768 // first unroll part.
2769 Value *StartVal = (part == 0) ? VectorStart : Identity;
2770 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2771 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2772 LoopVectorBody.back());
2775 // Before each round, move the insertion point right between
2776 // the PHIs and the values we are going to write.
2777 // This allows us to write both PHINodes and the extractelement
2779 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2781 VectorParts RdxParts;
2782 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2783 for (unsigned part = 0; part < UF; ++part) {
2784 // This PHINode contains the vectorized reduction variable, or
2785 // the initial value vector, if we bypass the vector loop.
2786 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2787 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2788 Value *StartVal = (part == 0) ? VectorStart : Identity;
2789 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2790 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2791 NewPhi->addIncoming(RdxExitVal[part],
2792 LoopVectorBody.back());
2793 RdxParts.push_back(NewPhi);
2796 // Reduce all of the unrolled parts into a single vector.
2797 Value *ReducedPartRdx = RdxParts[0];
2798 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2799 setDebugLocFromInst(Builder, ReducedPartRdx);
2800 for (unsigned part = 1; part < UF; ++part) {
2801 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2802 // Floating point operations had to be 'fast' to enable the reduction.
2803 ReducedPartRdx = addFastMathFlag(
2804 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2805 ReducedPartRdx, "bin.rdx"));
2807 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2808 ReducedPartRdx, RdxParts[part]);
2812 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2813 // and vector ops, reducing the set of values being computed by half each
2815 assert(isPowerOf2_32(VF) &&
2816 "Reduction emission only supported for pow2 vectors!");
2817 Value *TmpVec = ReducedPartRdx;
2818 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2819 for (unsigned i = VF; i != 1; i >>= 1) {
2820 // Move the upper half of the vector to the lower half.
2821 for (unsigned j = 0; j != i/2; ++j)
2822 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2824 // Fill the rest of the mask with undef.
2825 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2826 UndefValue::get(Builder.getInt32Ty()));
2829 Builder.CreateShuffleVector(TmpVec,
2830 UndefValue::get(TmpVec->getType()),
2831 ConstantVector::get(ShuffleMask),
2834 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2835 // Floating point operations had to be 'fast' to enable the reduction.
2836 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2837 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2839 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2842 // The result is in the first element of the vector.
2843 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2844 Builder.getInt32(0));
2847 // Create a phi node that merges control-flow from the backedge-taken check
2848 // block and the middle block.
2849 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
2850 LoopScalarPreHeader->getTerminator());
2851 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
2852 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2854 // Now, we need to fix the users of the reduction variable
2855 // inside and outside of the scalar remainder loop.
2856 // We know that the loop is in LCSSA form. We need to update the
2857 // PHI nodes in the exit blocks.
2858 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2859 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2860 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2861 if (!LCSSAPhi) break;
2863 // All PHINodes need to have a single entry edge, or two if
2864 // we already fixed them.
2865 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2867 // We found our reduction value exit-PHI. Update it with the
2868 // incoming bypass edge.
2869 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2870 // Add an edge coming from the bypass.
2871 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2874 }// end of the LCSSA phi scan.
2876 // Fix the scalar loop reduction variable with the incoming reduction sum
2877 // from the vector body and from the backedge value.
2878 int IncomingEdgeBlockIdx =
2879 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2880 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2881 // Pick the other block.
2882 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2883 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
2884 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2885 }// end of for each redux variable.
2889 // Remove redundant induction instructions.
2890 cse(LoopVectorBody);
2893 void InnerLoopVectorizer::fixLCSSAPHIs() {
2894 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2895 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2896 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2897 if (!LCSSAPhi) break;
2898 if (LCSSAPhi->getNumIncomingValues() == 1)
2899 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2904 InnerLoopVectorizer::VectorParts
2905 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2906 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2909 // Look for cached value.
2910 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2911 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2912 if (ECEntryIt != MaskCache.end())
2913 return ECEntryIt->second;
2915 VectorParts SrcMask = createBlockInMask(Src);
2917 // The terminator has to be a branch inst!
2918 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2919 assert(BI && "Unexpected terminator found");
2921 if (BI->isConditional()) {
2922 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2924 if (BI->getSuccessor(0) != Dst)
2925 for (unsigned part = 0; part < UF; ++part)
2926 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2928 for (unsigned part = 0; part < UF; ++part)
2929 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2931 MaskCache[Edge] = EdgeMask;
2935 MaskCache[Edge] = SrcMask;
2939 InnerLoopVectorizer::VectorParts
2940 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2941 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2943 // Loop incoming mask is all-one.
2944 if (OrigLoop->getHeader() == BB) {
2945 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2946 return getVectorValue(C);
2949 // This is the block mask. We OR all incoming edges, and with zero.
2950 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2951 VectorParts BlockMask = getVectorValue(Zero);
2954 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2955 VectorParts EM = createEdgeMask(*it, BB);
2956 for (unsigned part = 0; part < UF; ++part)
2957 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2963 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2964 InnerLoopVectorizer::VectorParts &Entry,
2965 unsigned UF, unsigned VF, PhiVector *PV) {
2966 PHINode* P = cast<PHINode>(PN);
2967 // Handle reduction variables:
2968 if (Legal->getReductionVars()->count(P)) {
2969 for (unsigned part = 0; part < UF; ++part) {
2970 // This is phase one of vectorizing PHIs.
2971 Type *VecTy = (VF == 1) ? PN->getType() :
2972 VectorType::get(PN->getType(), VF);
2973 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2974 LoopVectorBody.back()-> getFirstInsertionPt());
2980 setDebugLocFromInst(Builder, P);
2981 // Check for PHI nodes that are lowered to vector selects.
2982 if (P->getParent() != OrigLoop->getHeader()) {
2983 // We know that all PHIs in non-header blocks are converted into
2984 // selects, so we don't have to worry about the insertion order and we
2985 // can just use the builder.
2986 // At this point we generate the predication tree. There may be
2987 // duplications since this is a simple recursive scan, but future
2988 // optimizations will clean it up.
2990 unsigned NumIncoming = P->getNumIncomingValues();
2992 // Generate a sequence of selects of the form:
2993 // SELECT(Mask3, In3,
2994 // SELECT(Mask2, In2,
2996 for (unsigned In = 0; In < NumIncoming; In++) {
2997 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2999 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3001 for (unsigned part = 0; part < UF; ++part) {
3002 // We might have single edge PHIs (blocks) - use an identity
3003 // 'select' for the first PHI operand.
3005 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3008 // Select between the current value and the previous incoming edge
3009 // based on the incoming mask.
3010 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3011 Entry[part], "predphi");
3017 // This PHINode must be an induction variable.
3018 // Make sure that we know about it.
3019 assert(Legal->getInductionVars()->count(P) &&
3020 "Not an induction variable");
3022 LoopVectorizationLegality::InductionInfo II =
3023 Legal->getInductionVars()->lookup(P);
3026 case LoopVectorizationLegality::IK_NoInduction:
3027 llvm_unreachable("Unknown induction");
3028 case LoopVectorizationLegality::IK_IntInduction: {
3029 assert(P->getType() == II.StartValue->getType() && "Types must match");
3030 Type *PhiTy = P->getType();
3032 if (P == OldInduction) {
3033 // Handle the canonical induction variable. We might have had to
3035 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
3037 // Handle other induction variables that are now based on the
3039 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
3041 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
3042 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
3045 Broadcasted = getBroadcastInstrs(Broadcasted);
3046 // After broadcasting the induction variable we need to make the vector
3047 // consecutive by adding 0, 1, 2, etc.
3048 for (unsigned part = 0; part < UF; ++part)
3049 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
3052 case LoopVectorizationLegality::IK_ReverseIntInduction:
3053 case LoopVectorizationLegality::IK_PtrInduction:
3054 case LoopVectorizationLegality::IK_ReversePtrInduction:
3055 // Handle reverse integer and pointer inductions.
3056 Value *StartIdx = ExtendedIdx;
3057 // This is the normalized GEP that starts counting at zero.
3058 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
3061 // Handle the reverse integer induction variable case.
3062 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
3063 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
3064 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
3066 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
3069 // This is a new value so do not hoist it out.
3070 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
3071 // After broadcasting the induction variable we need to make the
3072 // vector consecutive by adding ... -3, -2, -1, 0.
3073 for (unsigned part = 0; part < UF; ++part)
3074 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
3079 // Handle the pointer induction variable case.
3080 assert(P->getType()->isPointerTy() && "Unexpected type.");
3082 // Is this a reverse induction ptr or a consecutive induction ptr.
3083 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
3086 // This is the vector of results. Notice that we don't generate
3087 // vector geps because scalar geps result in better code.
3088 for (unsigned part = 0; part < UF; ++part) {
3090 int EltIndex = (part) * (Reverse ? -1 : 1);
3091 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3094 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3096 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3098 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3100 Entry[part] = SclrGep;
3104 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3105 for (unsigned int i = 0; i < VF; ++i) {
3106 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
3107 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3110 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3112 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3114 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3116 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3117 Builder.getInt32(i),
3120 Entry[part] = VecVal;
3126 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3127 // For each instruction in the old loop.
3128 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3129 VectorParts &Entry = WidenMap.get(it);
3130 switch (it->getOpcode()) {
3131 case Instruction::Br:
3132 // Nothing to do for PHIs and BR, since we already took care of the
3133 // loop control flow instructions.
3135 case Instruction::PHI:{
3136 // Vectorize PHINodes.
3137 widenPHIInstruction(it, Entry, UF, VF, PV);
3141 case Instruction::Add:
3142 case Instruction::FAdd:
3143 case Instruction::Sub:
3144 case Instruction::FSub:
3145 case Instruction::Mul:
3146 case Instruction::FMul:
3147 case Instruction::UDiv:
3148 case Instruction::SDiv:
3149 case Instruction::FDiv:
3150 case Instruction::URem:
3151 case Instruction::SRem:
3152 case Instruction::FRem:
3153 case Instruction::Shl:
3154 case Instruction::LShr:
3155 case Instruction::AShr:
3156 case Instruction::And:
3157 case Instruction::Or:
3158 case Instruction::Xor: {
3159 // Just widen binops.
3160 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3161 setDebugLocFromInst(Builder, BinOp);
3162 VectorParts &A = getVectorValue(it->getOperand(0));
3163 VectorParts &B = getVectorValue(it->getOperand(1));
3165 // Use this vector value for all users of the original instruction.
3166 for (unsigned Part = 0; Part < UF; ++Part) {
3167 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3169 // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
3170 BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
3171 if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
3172 VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
3173 VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
3175 if (VecOp && isa<PossiblyExactOperator>(VecOp))
3176 VecOp->setIsExact(BinOp->isExact());
3178 // Copy the fast-math flags.
3179 if (VecOp && isa<FPMathOperator>(V))
3180 VecOp->setFastMathFlags(it->getFastMathFlags());
3185 propagateMetadata(Entry, it);
3188 case Instruction::Select: {
3190 // If the selector is loop invariant we can create a select
3191 // instruction with a scalar condition. Otherwise, use vector-select.
3192 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3194 setDebugLocFromInst(Builder, it);
3196 // The condition can be loop invariant but still defined inside the
3197 // loop. This means that we can't just use the original 'cond' value.
3198 // We have to take the 'vectorized' value and pick the first lane.
3199 // Instcombine will make this a no-op.
3200 VectorParts &Cond = getVectorValue(it->getOperand(0));
3201 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3202 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3204 Value *ScalarCond = (VF == 1) ? Cond[0] :
3205 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3207 for (unsigned Part = 0; Part < UF; ++Part) {
3208 Entry[Part] = Builder.CreateSelect(
3209 InvariantCond ? ScalarCond : Cond[Part],
3214 propagateMetadata(Entry, it);
3218 case Instruction::ICmp:
3219 case Instruction::FCmp: {
3220 // Widen compares. Generate vector compares.
3221 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3222 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3223 setDebugLocFromInst(Builder, it);
3224 VectorParts &A = getVectorValue(it->getOperand(0));
3225 VectorParts &B = getVectorValue(it->getOperand(1));
3226 for (unsigned Part = 0; Part < UF; ++Part) {
3229 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3231 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3235 propagateMetadata(Entry, it);
3239 case Instruction::Store:
3240 case Instruction::Load:
3241 vectorizeMemoryInstruction(it);
3243 case Instruction::ZExt:
3244 case Instruction::SExt:
3245 case Instruction::FPToUI:
3246 case Instruction::FPToSI:
3247 case Instruction::FPExt:
3248 case Instruction::PtrToInt:
3249 case Instruction::IntToPtr:
3250 case Instruction::SIToFP:
3251 case Instruction::UIToFP:
3252 case Instruction::Trunc:
3253 case Instruction::FPTrunc:
3254 case Instruction::BitCast: {
3255 CastInst *CI = dyn_cast<CastInst>(it);
3256 setDebugLocFromInst(Builder, it);
3257 /// Optimize the special case where the source is the induction
3258 /// variable. Notice that we can only optimize the 'trunc' case
3259 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3260 /// c. other casts depend on pointer size.
3261 if (CI->getOperand(0) == OldInduction &&
3262 it->getOpcode() == Instruction::Trunc) {
3263 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3265 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3266 for (unsigned Part = 0; Part < UF; ++Part)
3267 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
3268 propagateMetadata(Entry, it);
3271 /// Vectorize casts.
3272 Type *DestTy = (VF == 1) ? CI->getType() :
3273 VectorType::get(CI->getType(), VF);
3275 VectorParts &A = getVectorValue(it->getOperand(0));
3276 for (unsigned Part = 0; Part < UF; ++Part)
3277 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3278 propagateMetadata(Entry, it);
3282 case Instruction::Call: {
3283 // Ignore dbg intrinsics.
3284 if (isa<DbgInfoIntrinsic>(it))
3286 setDebugLocFromInst(Builder, it);
3288 Module *M = BB->getParent()->getParent();
3289 CallInst *CI = cast<CallInst>(it);
3290 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3291 assert(ID && "Not an intrinsic call!");
3293 case Intrinsic::lifetime_end:
3294 case Intrinsic::lifetime_start:
3295 scalarizeInstruction(it);
3298 bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
3299 for (unsigned Part = 0; Part < UF; ++Part) {
3300 SmallVector<Value *, 4> Args;
3301 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3302 if (HasScalarOpd && i == 1) {
3303 Args.push_back(CI->getArgOperand(i));
3306 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3307 Args.push_back(Arg[Part]);
3309 Type *Tys[] = {CI->getType()};
3311 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3313 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3314 Entry[Part] = Builder.CreateCall(F, Args);
3317 propagateMetadata(Entry, it);
3324 // All other instructions are unsupported. Scalarize them.
3325 scalarizeInstruction(it);
3328 }// end of for_each instr.
3331 void InnerLoopVectorizer::updateAnalysis() {
3332 // Forget the original basic block.
3333 SE->forgetLoop(OrigLoop);
3335 // Update the dominator tree information.
3336 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3337 "Entry does not dominate exit.");
3339 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3340 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3341 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3343 // Due to if predication of stores we might create a sequence of "if(pred)
3344 // a[i] = ...; " blocks.
3345 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3347 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3348 else if (isPredicatedBlock(i)) {
3349 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3351 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3355 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3356 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3357 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3358 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
3360 DEBUG(DT->verifyDomTree());
3363 /// \brief Check whether it is safe to if-convert this phi node.
3365 /// Phi nodes with constant expressions that can trap are not safe to if
3367 static bool canIfConvertPHINodes(BasicBlock *BB) {
3368 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3369 PHINode *Phi = dyn_cast<PHINode>(I);
3372 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3373 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3380 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3381 if (!EnableIfConversion) {
3382 emitAnalysis(Report() << "if-conversion is disabled");
3386 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3388 // A list of pointers that we can safely read and write to.
3389 SmallPtrSet<Value *, 8> SafePointes;
3391 // Collect safe addresses.
3392 for (Loop::block_iterator BI = TheLoop->block_begin(),
3393 BE = TheLoop->block_end(); BI != BE; ++BI) {
3394 BasicBlock *BB = *BI;
3396 if (blockNeedsPredication(BB))
3399 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3400 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3401 SafePointes.insert(LI->getPointerOperand());
3402 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3403 SafePointes.insert(SI->getPointerOperand());
3407 // Collect the blocks that need predication.
3408 BasicBlock *Header = TheLoop->getHeader();
3409 for (Loop::block_iterator BI = TheLoop->block_begin(),
3410 BE = TheLoop->block_end(); BI != BE; ++BI) {
3411 BasicBlock *BB = *BI;
3413 // We don't support switch statements inside loops.
3414 if (!isa<BranchInst>(BB->getTerminator())) {
3415 emitAnalysis(Report(BB->getTerminator())
3416 << "loop contains a switch statement");
3420 // We must be able to predicate all blocks that need to be predicated.
3421 if (blockNeedsPredication(BB)) {
3422 if (!blockCanBePredicated(BB, SafePointes)) {
3423 emitAnalysis(Report(BB->getTerminator())
3424 << "control flow cannot be substituted for a select");
3427 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3428 emitAnalysis(Report(BB->getTerminator())
3429 << "control flow cannot be substituted for a select");
3434 // We can if-convert this loop.
3438 bool LoopVectorizationLegality::canVectorize() {
3439 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3440 // be canonicalized.
3441 if (!TheLoop->getLoopPreheader()) {
3443 Report() << "loop control flow is not understood by vectorizer");
3447 // We can only vectorize innermost loops.
3448 if (TheLoop->getSubLoopsVector().size()) {
3449 emitAnalysis(Report() << "loop is not the innermost loop");
3453 // We must have a single backedge.
3454 if (TheLoop->getNumBackEdges() != 1) {
3456 Report() << "loop control flow is not understood by vectorizer");
3460 // We must have a single exiting block.
3461 if (!TheLoop->getExitingBlock()) {
3463 Report() << "loop control flow is not understood by vectorizer");
3467 // We need to have a loop header.
3468 DEBUG(dbgs() << "LV: Found a loop: " <<
3469 TheLoop->getHeader()->getName() << '\n');
3471 // Check if we can if-convert non-single-bb loops.
3472 unsigned NumBlocks = TheLoop->getNumBlocks();
3473 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3474 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3478 // ScalarEvolution needs to be able to find the exit count.
3479 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3480 if (ExitCount == SE->getCouldNotCompute()) {
3481 emitAnalysis(Report() << "could not determine number of loop iterations");
3482 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3486 // Check if we can vectorize the instructions and CFG in this loop.
3487 if (!canVectorizeInstrs()) {
3488 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3492 // Go over each instruction and look at memory deps.
3493 if (!canVectorizeMemory()) {
3494 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3498 // Collect all of the variables that remain uniform after vectorization.
3499 collectLoopUniforms();
3501 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3502 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
3505 // Okay! We can vectorize. At this point we don't have any other mem analysis
3506 // which may limit our maximum vectorization factor, so just return true with
3511 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3512 if (Ty->isPointerTy())
3513 return DL.getIntPtrType(Ty);
3515 // It is possible that char's or short's overflow when we ask for the loop's
3516 // trip count, work around this by changing the type size.
3517 if (Ty->getScalarSizeInBits() < 32)
3518 return Type::getInt32Ty(Ty->getContext());
3523 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3524 Ty0 = convertPointerToIntegerType(DL, Ty0);
3525 Ty1 = convertPointerToIntegerType(DL, Ty1);
3526 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3531 /// \brief Check that the instruction has outside loop users and is not an
3532 /// identified reduction variable.
3533 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3534 SmallPtrSet<Value *, 4> &Reductions) {
3535 // Reduction instructions are allowed to have exit users. All other
3536 // instructions must not have external users.
3537 if (!Reductions.count(Inst))
3538 //Check that all of the users of the loop are inside the BB.
3539 for (User *U : Inst->users()) {
3540 Instruction *UI = cast<Instruction>(U);
3541 // This user may be a reduction exit value.
3542 if (!TheLoop->contains(UI)) {
3543 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3550 bool LoopVectorizationLegality::canVectorizeInstrs() {
3551 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3552 BasicBlock *Header = TheLoop->getHeader();
3554 // Look for the attribute signaling the absence of NaNs.
3555 Function &F = *Header->getParent();
3556 if (F.hasFnAttribute("no-nans-fp-math"))
3557 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3558 AttributeSet::FunctionIndex,
3559 "no-nans-fp-math").getValueAsString() == "true";
3561 // For each block in the loop.
3562 for (Loop::block_iterator bb = TheLoop->block_begin(),
3563 be = TheLoop->block_end(); bb != be; ++bb) {
3565 // Scan the instructions in the block and look for hazards.
3566 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3569 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3570 Type *PhiTy = Phi->getType();
3571 // Check that this PHI type is allowed.
3572 if (!PhiTy->isIntegerTy() &&
3573 !PhiTy->isFloatingPointTy() &&
3574 !PhiTy->isPointerTy()) {
3575 emitAnalysis(Report(it)
3576 << "loop control flow is not understood by vectorizer");
3577 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3581 // If this PHINode is not in the header block, then we know that we
3582 // can convert it to select during if-conversion. No need to check if
3583 // the PHIs in this block are induction or reduction variables.
3584 if (*bb != Header) {
3585 // Check that this instruction has no outside users or is an
3586 // identified reduction value with an outside user.
3587 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3589 emitAnalysis(Report(it) << "value could not be identified as "
3590 "an induction or reduction variable");
3594 // We only allow if-converted PHIs with more than two incoming values.
3595 if (Phi->getNumIncomingValues() != 2) {
3596 emitAnalysis(Report(it)
3597 << "control flow not understood by vectorizer");
3598 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3602 // This is the value coming from the preheader.
3603 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3604 // Check if this is an induction variable.
3605 InductionKind IK = isInductionVariable(Phi);
3607 if (IK_NoInduction != IK) {
3608 // Get the widest type.
3610 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3612 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3614 // Int inductions are special because we only allow one IV.
3615 if (IK == IK_IntInduction) {
3616 // Use the phi node with the widest type as induction. Use the last
3617 // one if there are multiple (no good reason for doing this other
3618 // than it is expedient).
3619 if (!Induction || PhiTy == WidestIndTy)
3623 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3624 Inductions[Phi] = InductionInfo(StartValue, IK);
3626 // Until we explicitly handle the case of an induction variable with
3627 // an outside loop user we have to give up vectorizing this loop.
3628 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3629 emitAnalysis(Report(it) << "use of induction value outside of the "
3630 "loop is not handled by vectorizer");
3637 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3638 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3641 if (AddReductionVar(Phi, RK_IntegerMult)) {
3642 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3645 if (AddReductionVar(Phi, RK_IntegerOr)) {
3646 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3649 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3650 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3653 if (AddReductionVar(Phi, RK_IntegerXor)) {
3654 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3657 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3658 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3661 if (AddReductionVar(Phi, RK_FloatMult)) {
3662 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3665 if (AddReductionVar(Phi, RK_FloatAdd)) {
3666 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3669 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3670 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3675 emitAnalysis(Report(it) << "value that could not be identified as "
3676 "reduction is used outside the loop");
3677 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3679 }// end of PHI handling
3681 // We still don't handle functions. However, we can ignore dbg intrinsic
3682 // calls and we do handle certain intrinsic and libm functions.
3683 CallInst *CI = dyn_cast<CallInst>(it);
3684 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3685 emitAnalysis(Report(it) << "call instruction cannot be vectorized");
3686 DEBUG(dbgs() << "LV: Found a call site.\n");
3690 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3691 // second argument is the same (i.e. loop invariant)
3693 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3694 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3695 emitAnalysis(Report(it)
3696 << "intrinsic instruction cannot be vectorized");
3697 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3702 // Check that the instruction return type is vectorizable.
3703 // Also, we can't vectorize extractelement instructions.
3704 if ((!VectorType::isValidElementType(it->getType()) &&
3705 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3706 emitAnalysis(Report(it)
3707 << "instruction return type cannot be vectorized");
3708 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3712 // Check that the stored type is vectorizable.
3713 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3714 Type *T = ST->getValueOperand()->getType();
3715 if (!VectorType::isValidElementType(T)) {
3716 emitAnalysis(Report(ST) << "store instruction cannot be vectorized");
3719 if (EnableMemAccessVersioning)
3720 collectStridedAcccess(ST);
3723 if (EnableMemAccessVersioning)
3724 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3725 collectStridedAcccess(LI);
3727 // Reduction instructions are allowed to have exit users.
3728 // All other instructions must not have external users.
3729 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3730 emitAnalysis(Report(it) << "value cannot be used outside the loop");
3739 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3740 if (Inductions.empty()) {
3741 emitAnalysis(Report()
3742 << "loop induction variable could not be identified");
3750 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3751 /// return the induction operand of the gep pointer.
3752 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3753 const DataLayout *DL, Loop *Lp) {
3754 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3758 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3760 // Check that all of the gep indices are uniform except for our induction
3762 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3763 if (i != InductionOperand &&
3764 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3766 return GEP->getOperand(InductionOperand);
3769 ///\brief Look for a cast use of the passed value.
3770 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3771 Value *UniqueCast = nullptr;
3772 for (User *U : Ptr->users()) {
3773 CastInst *CI = dyn_cast<CastInst>(U);
3774 if (CI && CI->getType() == Ty) {
3784 ///\brief Get the stride of a pointer access in a loop.
3785 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3786 /// pointer to the Value, or null otherwise.
3787 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3788 const DataLayout *DL, Loop *Lp) {
3789 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3790 if (!PtrTy || PtrTy->isAggregateType())
3793 // Try to remove a gep instruction to make the pointer (actually index at this
3794 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3795 // pointer, otherwise, we are analyzing the index.
3796 Value *OrigPtr = Ptr;
3798 // The size of the pointer access.
3799 int64_t PtrAccessSize = 1;
3801 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3802 const SCEV *V = SE->getSCEV(Ptr);
3806 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3807 V = C->getOperand();
3809 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3813 V = S->getStepRecurrence(*SE);
3817 // Strip off the size of access multiplication if we are still analyzing the
3819 if (OrigPtr == Ptr) {
3820 DL->getTypeAllocSize(PtrTy->getElementType());
3821 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3822 if (M->getOperand(0)->getSCEVType() != scConstant)
3825 const APInt &APStepVal =
3826 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3828 // Huge step value - give up.
3829 if (APStepVal.getBitWidth() > 64)
3832 int64_t StepVal = APStepVal.getSExtValue();
3833 if (PtrAccessSize != StepVal)
3835 V = M->getOperand(1);
3840 Type *StripedOffRecurrenceCast = nullptr;
3841 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3842 StripedOffRecurrenceCast = C->getType();
3843 V = C->getOperand();
3846 // Look for the loop invariant symbolic value.
3847 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3851 Value *Stride = U->getValue();
3852 if (!Lp->isLoopInvariant(Stride))
3855 // If we have stripped off the recurrence cast we have to make sure that we
3856 // return the value that is used in this loop so that we can replace it later.
3857 if (StripedOffRecurrenceCast)
3858 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3863 void LoopVectorizationLegality::collectStridedAcccess(Value *MemAccess) {
3864 Value *Ptr = nullptr;
3865 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3866 Ptr = LI->getPointerOperand();
3867 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3868 Ptr = SI->getPointerOperand();
3872 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3876 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3877 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3878 Strides[Ptr] = Stride;
3879 StrideSet.insert(Stride);
3882 void LoopVectorizationLegality::collectLoopUniforms() {
3883 // We now know that the loop is vectorizable!
3884 // Collect variables that will remain uniform after vectorization.
3885 std::vector<Value*> Worklist;
3886 BasicBlock *Latch = TheLoop->getLoopLatch();
3888 // Start with the conditional branch and walk up the block.
3889 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3891 // Also add all consecutive pointer values; these values will be uniform
3892 // after vectorization (and subsequent cleanup) and, until revectorization is
3893 // supported, all dependencies must also be uniform.
3894 for (Loop::block_iterator B = TheLoop->block_begin(),
3895 BE = TheLoop->block_end(); B != BE; ++B)
3896 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
3898 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
3899 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3901 while (Worklist.size()) {
3902 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3903 Worklist.pop_back();
3905 // Look at instructions inside this loop.
3906 // Stop when reaching PHI nodes.
3907 // TODO: we need to follow values all over the loop, not only in this block.
3908 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3911 // This is a known uniform.
3914 // Insert all operands.
3915 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3920 /// \brief Analyses memory accesses in a loop.
3922 /// Checks whether run time pointer checks are needed and builds sets for data
3923 /// dependence checking.
3924 class AccessAnalysis {
3926 /// \brief Read or write access location.
3927 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3928 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
3930 /// \brief Set of potential dependent memory accesses.
3931 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
3933 AccessAnalysis(const DataLayout *Dl, AliasAnalysis *AA, DepCandidates &DA) :
3934 DL(Dl), AST(*AA), DepCands(DA), IsRTCheckNeeded(false) {}
3936 /// \brief Register a load and whether it is only read from.
3937 void addLoad(AliasAnalysis::Location &Loc, bool IsReadOnly) {
3938 Value *Ptr = const_cast<Value*>(Loc.Ptr);
3939 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
3940 Accesses.insert(MemAccessInfo(Ptr, false));
3942 ReadOnlyPtr.insert(Ptr);
3945 /// \brief Register a store.
3946 void addStore(AliasAnalysis::Location &Loc) {
3947 Value *Ptr = const_cast<Value*>(Loc.Ptr);
3948 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
3949 Accesses.insert(MemAccessInfo(Ptr, true));
3952 /// \brief Check whether we can check the pointers at runtime for
3953 /// non-intersection.
3954 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
3955 unsigned &NumComparisons, ScalarEvolution *SE,
3956 Loop *TheLoop, ValueToValueMap &Strides,
3957 bool ShouldCheckStride = false);
3959 /// \brief Goes over all memory accesses, checks whether a RT check is needed
3960 /// and builds sets of dependent accesses.
3961 void buildDependenceSets() {
3962 processMemAccesses();
3965 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
3967 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
3968 void resetDepChecks() { CheckDeps.clear(); }
3970 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
3973 typedef SetVector<MemAccessInfo> PtrAccessSet;
3975 /// \brief Go over all memory access and check whether runtime pointer checks
3976 /// are needed /// and build sets of dependency check candidates.
3977 void processMemAccesses();
3979 /// Set of all accesses.
3980 PtrAccessSet Accesses;
3982 /// Set of accesses that need a further dependence check.
3983 MemAccessInfoSet CheckDeps;
3985 /// Set of pointers that are read only.
3986 SmallPtrSet<Value*, 16> ReadOnlyPtr;
3988 const DataLayout *DL;
3990 /// An alias set tracker to partition the access set by underlying object and
3991 //intrinsic property (such as TBAA metadata).
3992 AliasSetTracker AST;
3994 /// Sets of potentially dependent accesses - members of one set share an
3995 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
3996 /// dependence check.
3997 DepCandidates &DepCands;
3999 bool IsRTCheckNeeded;
4002 } // end anonymous namespace
4004 /// \brief Check whether a pointer can participate in a runtime bounds check.
4005 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
4007 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
4008 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4012 return AR->isAffine();
4015 /// \brief Check the stride of the pointer and ensure that it does not wrap in
4016 /// the address space.
4017 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4018 const Loop *Lp, ValueToValueMap &StridesMap);
4020 bool AccessAnalysis::canCheckPtrAtRT(
4021 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
4022 unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
4023 ValueToValueMap &StridesMap, bool ShouldCheckStride) {
4024 // Find pointers with computable bounds. We are going to use this information
4025 // to place a runtime bound check.
4026 bool CanDoRT = true;
4028 bool IsDepCheckNeeded = isDependencyCheckNeeded();
4031 // We assign a consecutive id to access from different alias sets.
4032 // Accesses between different groups doesn't need to be checked.
4034 for (auto &AS : AST) {
4035 unsigned NumReadPtrChecks = 0;
4036 unsigned NumWritePtrChecks = 0;
4038 // We assign consecutive id to access from different dependence sets.
4039 // Accesses within the same set don't need a runtime check.
4040 unsigned RunningDepId = 1;
4041 DenseMap<Value *, unsigned> DepSetId;
4044 Value *Ptr = A.getValue();
4045 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
4046 MemAccessInfo Access(Ptr, IsWrite);
4049 ++NumWritePtrChecks;
4053 if (hasComputableBounds(SE, StridesMap, Ptr) &&
4054 // When we run after a failing dependency check we have to make sure we
4055 // don't have wrapping pointers.
4056 (!ShouldCheckStride ||
4057 isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
4058 // The id of the dependence set.
4061 if (IsDepCheckNeeded) {
4062 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
4063 unsigned &LeaderId = DepSetId[Leader];
4065 LeaderId = RunningDepId++;
4068 // Each access has its own dependence set.
4069 DepId = RunningDepId++;
4071 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
4073 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
4079 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
4080 NumComparisons += 0; // Only one dependence set.
4082 NumComparisons += (NumWritePtrChecks * (NumReadPtrChecks +
4083 NumWritePtrChecks - 1));
4089 // If the pointers that we would use for the bounds comparison have different
4090 // address spaces, assume the values aren't directly comparable, so we can't
4091 // use them for the runtime check. We also have to assume they could
4092 // overlap. In the future there should be metadata for whether address spaces
4094 unsigned NumPointers = RtCheck.Pointers.size();
4095 for (unsigned i = 0; i < NumPointers; ++i) {
4096 for (unsigned j = i + 1; j < NumPointers; ++j) {
4097 // Only need to check pointers between two different dependency sets.
4098 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
4100 // Only need to check pointers in the same alias set.
4101 if (RtCheck.AliasSetId[i] != RtCheck.AliasSetId[j])
4104 Value *PtrI = RtCheck.Pointers[i];
4105 Value *PtrJ = RtCheck.Pointers[j];
4107 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
4108 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
4110 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
4111 " different address spaces\n");
4120 void AccessAnalysis::processMemAccesses() {
4121 // We process the set twice: first we process read-write pointers, last we
4122 // process read-only pointers. This allows us to skip dependence tests for
4123 // read-only pointers.
4125 DEBUG(dbgs() << "LV: Processing memory accesses...\n");
4126 DEBUG(dbgs() << " AST: "; AST.dump());
4127 DEBUG(dbgs() << "LV: Accesses:\n");
4129 for (auto A : Accesses)
4130 dbgs() << "\t" << *A.getPointer() << " (" <<
4131 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
4132 "read-only" : "read")) << ")\n";
4135 // The AliasSetTracker has nicely partitioned our pointers by metadata
4136 // compatibility and potential for underlying-object overlap. As a result, we
4137 // only need to check for potential pointer dependencies within each alias
4139 for (auto &AS : AST) {
4140 // Note that both the alias-set tracker and the alias sets themselves used
4141 // linked lists internally and so the iteration order here is deterministic
4142 // (matching the original instruction order within each set).
4144 bool SetHasWrite = false;
4146 // Map of pointers to last access encountered.
4147 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
4148 UnderlyingObjToAccessMap ObjToLastAccess;
4150 // Set of access to check after all writes have been processed.
4151 PtrAccessSet DeferredAccesses;
4153 // Iterate over each alias set twice, once to process read/write pointers,
4154 // and then to process read-only pointers.
4155 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
4156 bool UseDeferred = SetIteration > 0;
4157 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
4160 Value *Ptr = A.getValue();
4161 bool IsWrite = S.count(MemAccessInfo(Ptr, true));
4163 // If we're using the deferred access set, then it contains only reads.
4164 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
4165 if (UseDeferred && !IsReadOnlyPtr)
4167 // Otherwise, the pointer must be in the PtrAccessSet, either as a read
4169 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
4170 S.count(MemAccessInfo(Ptr, false))) &&
4171 "Alias-set pointer not in the access set?");
4173 MemAccessInfo Access(Ptr, IsWrite);
4174 DepCands.insert(Access);
4176 // Memorize read-only pointers for later processing and skip them in the
4177 // first round (they need to be checked after we have seen all write
4178 // pointers). Note: we also mark pointer that are not consecutive as
4179 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need
4180 // the second check for "!IsWrite".
4181 if (!UseDeferred && IsReadOnlyPtr) {
4182 DeferredAccesses.insert(Access);
4186 // If this is a write - check other reads and writes for conflicts. If
4187 // this is a read only check other writes for conflicts (but only if
4188 // there is no other write to the ptr - this is an optimization to
4189 // catch "a[i] = a[i] + " without having to do a dependence check).
4190 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
4191 CheckDeps.insert(Access);
4192 IsRTCheckNeeded = true;
4198 // Create sets of pointers connected by a shared alias set and
4199 // underlying object.
4200 typedef SmallVector<Value*, 16> ValueVector;
4201 ValueVector TempObjects;
4202 GetUnderlyingObjects(Ptr, TempObjects, DL);
4203 for (Value *UnderlyingObj : TempObjects) {
4204 UnderlyingObjToAccessMap::iterator Prev =
4205 ObjToLastAccess.find(UnderlyingObj);
4206 if (Prev != ObjToLastAccess.end())
4207 DepCands.unionSets(Access, Prev->second);
4209 ObjToLastAccess[UnderlyingObj] = Access;
4217 /// \brief Checks memory dependences among accesses to the same underlying
4218 /// object to determine whether there vectorization is legal or not (and at
4219 /// which vectorization factor).
4221 /// This class works under the assumption that we already checked that memory
4222 /// locations with different underlying pointers are "must-not alias".
4223 /// We use the ScalarEvolution framework to symbolically evalutate access
4224 /// functions pairs. Since we currently don't restructure the loop we can rely
4225 /// on the program order of memory accesses to determine their safety.
4226 /// At the moment we will only deem accesses as safe for:
4227 /// * A negative constant distance assuming program order.
4229 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
4230 /// a[i] = tmp; y = a[i];
4232 /// The latter case is safe because later checks guarantuee that there can't
4233 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
4234 /// the same variable: a header phi can only be an induction or a reduction, a
4235 /// reduction can't have a memory sink, an induction can't have a memory
4236 /// source). This is important and must not be violated (or we have to
4237 /// resort to checking for cycles through memory).
4239 /// * A positive constant distance assuming program order that is bigger
4240 /// than the biggest memory access.
4242 /// tmp = a[i] OR b[i] = x
4243 /// a[i+2] = tmp y = b[i+2];
4245 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
4247 /// * Zero distances and all accesses have the same size.
4249 class MemoryDepChecker {
4251 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
4252 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4254 MemoryDepChecker(ScalarEvolution *Se, const DataLayout *Dl, const Loop *L)
4255 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
4256 ShouldRetryWithRuntimeCheck(false) {}
4258 /// \brief Register the location (instructions are given increasing numbers)
4259 /// of a write access.
4260 void addAccess(StoreInst *SI) {
4261 Value *Ptr = SI->getPointerOperand();
4262 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
4263 InstMap.push_back(SI);
4267 /// \brief Register the location (instructions are given increasing numbers)
4268 /// of a write access.
4269 void addAccess(LoadInst *LI) {
4270 Value *Ptr = LI->getPointerOperand();
4271 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
4272 InstMap.push_back(LI);
4276 /// \brief Check whether the dependencies between the accesses are safe.
4278 /// Only checks sets with elements in \p CheckDeps.
4279 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4280 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
4282 /// \brief The maximum number of bytes of a vector register we can vectorize
4283 /// the accesses safely with.
4284 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
4286 /// \brief In same cases when the dependency check fails we can still
4287 /// vectorize the loop with a dynamic array access check.
4288 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
4291 ScalarEvolution *SE;
4292 const DataLayout *DL;
4293 const Loop *InnermostLoop;
4295 /// \brief Maps access locations (ptr, read/write) to program order.
4296 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
4298 /// \brief Memory access instructions in program order.
4299 SmallVector<Instruction *, 16> InstMap;
4301 /// \brief The program order index to be used for the next instruction.
4304 // We can access this many bytes in parallel safely.
4305 unsigned MaxSafeDepDistBytes;
4307 /// \brief If we see a non-constant dependence distance we can still try to
4308 /// vectorize this loop with runtime checks.
4309 bool ShouldRetryWithRuntimeCheck;
4311 /// \brief Check whether there is a plausible dependence between the two
4314 /// Access \p A must happen before \p B in program order. The two indices
4315 /// identify the index into the program order map.
4317 /// This function checks whether there is a plausible dependence (or the
4318 /// absence of such can't be proved) between the two accesses. If there is a
4319 /// plausible dependence but the dependence distance is bigger than one
4320 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
4321 /// distance is smaller than any other distance encountered so far).
4322 /// Otherwise, this function returns true signaling a possible dependence.
4323 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
4324 const MemAccessInfo &B, unsigned BIdx,
4325 ValueToValueMap &Strides);
4327 /// \brief Check whether the data dependence could prevent store-load
4329 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
4332 } // end anonymous namespace
4334 static bool isInBoundsGep(Value *Ptr) {
4335 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
4336 return GEP->isInBounds();
4340 /// \brief Check whether the access through \p Ptr has a constant stride.
4341 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4342 const Loop *Lp, ValueToValueMap &StridesMap) {
4343 const Type *Ty = Ptr->getType();
4344 assert(Ty->isPointerTy() && "Unexpected non-ptr");
4346 // Make sure that the pointer does not point to aggregate types.
4347 const PointerType *PtrTy = cast<PointerType>(Ty);
4348 if (PtrTy->getElementType()->isAggregateType()) {
4349 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
4354 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
4356 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4358 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
4359 << *Ptr << " SCEV: " << *PtrScev << "\n");
4363 // The accesss function must stride over the innermost loop.
4364 if (Lp != AR->getLoop()) {
4365 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
4366 *Ptr << " SCEV: " << *PtrScev << "\n");
4369 // The address calculation must not wrap. Otherwise, a dependence could be
4371 // An inbounds getelementptr that is a AddRec with a unit stride
4372 // cannot wrap per definition. The unit stride requirement is checked later.
4373 // An getelementptr without an inbounds attribute and unit stride would have
4374 // to access the pointer value "0" which is undefined behavior in address
4375 // space 0, therefore we can also vectorize this case.
4376 bool IsInBoundsGEP = isInBoundsGep(Ptr);
4377 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
4378 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
4379 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
4380 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
4381 << *Ptr << " SCEV: " << *PtrScev << "\n");
4385 // Check the step is constant.
4386 const SCEV *Step = AR->getStepRecurrence(*SE);
4388 // Calculate the pointer stride and check if it is consecutive.
4389 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4391 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
4392 " SCEV: " << *PtrScev << "\n");
4396 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
4397 const APInt &APStepVal = C->getValue()->getValue();
4399 // Huge step value - give up.
4400 if (APStepVal.getBitWidth() > 64)
4403 int64_t StepVal = APStepVal.getSExtValue();
4406 int64_t Stride = StepVal / Size;
4407 int64_t Rem = StepVal % Size;
4411 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
4412 // know we can't "wrap around the address space". In case of address space
4413 // zero we know that this won't happen without triggering undefined behavior.
4414 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
4415 Stride != 1 && Stride != -1)
4421 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
4422 unsigned TypeByteSize) {
4423 // If loads occur at a distance that is not a multiple of a feasible vector
4424 // factor store-load forwarding does not take place.
4425 // Positive dependences might cause troubles because vectorizing them might
4426 // prevent store-load forwarding making vectorized code run a lot slower.
4427 // a[i] = a[i-3] ^ a[i-8];
4428 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
4429 // hence on your typical architecture store-load forwarding does not take
4430 // place. Vectorizing in such cases does not make sense.
4431 // Store-load forwarding distance.
4432 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
4433 // Maximum vector factor.
4434 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
4435 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
4436 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
4438 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
4440 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
4441 MaxVFWithoutSLForwardIssues = (vf >>=1);
4446 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
4447 DEBUG(dbgs() << "LV: Distance " << Distance <<
4448 " that could cause a store-load forwarding conflict\n");
4452 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
4453 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
4454 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
4458 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
4459 const MemAccessInfo &B, unsigned BIdx,
4460 ValueToValueMap &Strides) {
4461 assert (AIdx < BIdx && "Must pass arguments in program order");
4463 Value *APtr = A.getPointer();
4464 Value *BPtr = B.getPointer();
4465 bool AIsWrite = A.getInt();
4466 bool BIsWrite = B.getInt();
4468 // Two reads are independent.
4469 if (!AIsWrite && !BIsWrite)
4472 // We cannot check pointers in different address spaces.
4473 if (APtr->getType()->getPointerAddressSpace() !=
4474 BPtr->getType()->getPointerAddressSpace())
4477 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
4478 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
4480 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
4481 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
4483 const SCEV *Src = AScev;
4484 const SCEV *Sink = BScev;
4486 // If the induction step is negative we have to invert source and sink of the
4488 if (StrideAPtr < 0) {
4491 std::swap(APtr, BPtr);
4492 std::swap(Src, Sink);
4493 std::swap(AIsWrite, BIsWrite);
4494 std::swap(AIdx, BIdx);
4495 std::swap(StrideAPtr, StrideBPtr);
4498 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
4500 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
4501 << "(Induction step: " << StrideAPtr << ")\n");
4502 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
4503 << *InstMap[BIdx] << ": " << *Dist << "\n");
4505 // Need consecutive accesses. We don't want to vectorize
4506 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
4507 // the address space.
4508 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
4509 DEBUG(dbgs() << "Non-consecutive pointer access\n");
4513 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
4515 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
4516 ShouldRetryWithRuntimeCheck = true;
4520 Type *ATy = APtr->getType()->getPointerElementType();
4521 Type *BTy = BPtr->getType()->getPointerElementType();
4522 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
4524 // Negative distances are not plausible dependencies.
4525 const APInt &Val = C->getValue()->getValue();
4526 if (Val.isNegative()) {
4527 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
4528 if (IsTrueDataDependence &&
4529 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
4533 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
4537 // Write to the same location with the same size.
4538 // Could be improved to assert type sizes are the same (i32 == float, etc).
4542 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
4546 assert(Val.isStrictlyPositive() && "Expect a positive value");
4548 // Positive distance bigger than max vectorization factor.
4551 "LV: ReadWrite-Write positive dependency with different types\n");
4555 unsigned Distance = (unsigned) Val.getZExtValue();
4557 // Bail out early if passed-in parameters make vectorization not feasible.
4558 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
4559 unsigned ForcedUnroll = VectorizationUnroll ? VectorizationUnroll : 1;
4561 // The distance must be bigger than the size needed for a vectorized version
4562 // of the operation and the size of the vectorized operation must not be
4563 // bigger than the currrent maximum size.
4564 if (Distance < 2*TypeByteSize ||
4565 2*TypeByteSize > MaxSafeDepDistBytes ||
4566 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
4567 DEBUG(dbgs() << "LV: Failure because of Positive distance "
4568 << Val.getSExtValue() << '\n');
4572 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
4573 Distance : MaxSafeDepDistBytes;
4575 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
4576 if (IsTrueDataDependence &&
4577 couldPreventStoreLoadForward(Distance, TypeByteSize))
4580 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
4581 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
4586 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4587 MemAccessInfoSet &CheckDeps,
4588 ValueToValueMap &Strides) {
4590 MaxSafeDepDistBytes = -1U;
4591 while (!CheckDeps.empty()) {
4592 MemAccessInfo CurAccess = *CheckDeps.begin();
4594 // Get the relevant memory access set.
4595 EquivalenceClasses<MemAccessInfo>::iterator I =
4596 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
4598 // Check accesses within this set.
4599 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
4600 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
4602 // Check every access pair.
4604 CheckDeps.erase(*AI);
4605 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
4607 // Check every accessing instruction pair in program order.
4608 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
4609 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
4610 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
4611 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
4612 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
4614 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
4625 bool LoopVectorizationLegality::canVectorizeMemory() {
4627 typedef SmallVector<Value*, 16> ValueVector;
4628 typedef SmallPtrSet<Value*, 16> ValueSet;
4630 // Holds the Load and Store *instructions*.
4634 // Holds all the different accesses in the loop.
4635 unsigned NumReads = 0;
4636 unsigned NumReadWrites = 0;
4638 PtrRtCheck.Pointers.clear();
4639 PtrRtCheck.Need = false;
4641 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
4642 MemoryDepChecker DepChecker(SE, DL, TheLoop);
4645 for (Loop::block_iterator bb = TheLoop->block_begin(),
4646 be = TheLoop->block_end(); bb != be; ++bb) {
4648 // Scan the BB and collect legal loads and stores.
4649 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4652 // If this is a load, save it. If this instruction can read from memory
4653 // but is not a load, then we quit. Notice that we don't handle function
4654 // calls that read or write.
4655 if (it->mayReadFromMemory()) {
4656 // Many math library functions read the rounding mode. We will only
4657 // vectorize a loop if it contains known function calls that don't set
4658 // the flag. Therefore, it is safe to ignore this read from memory.
4659 CallInst *Call = dyn_cast<CallInst>(it);
4660 if (Call && getIntrinsicIDForCall(Call, TLI))
4663 LoadInst *Ld = dyn_cast<LoadInst>(it);
4664 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
4665 emitAnalysis(Report(Ld)
4666 << "read with atomic ordering or volatile read");
4667 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
4671 Loads.push_back(Ld);
4672 DepChecker.addAccess(Ld);
4676 // Save 'store' instructions. Abort if other instructions write to memory.
4677 if (it->mayWriteToMemory()) {
4678 StoreInst *St = dyn_cast<StoreInst>(it);
4680 emitAnalysis(Report(it) << "instruction cannot be vectorized");
4683 if (!St->isSimple() && !IsAnnotatedParallel) {
4684 emitAnalysis(Report(St)
4685 << "write with atomic ordering or volatile write");
4686 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
4690 Stores.push_back(St);
4691 DepChecker.addAccess(St);
4696 // Now we have two lists that hold the loads and the stores.
4697 // Next, we find the pointers that they use.
4699 // Check if we see any stores. If there are no stores, then we don't
4700 // care if the pointers are *restrict*.
4701 if (!Stores.size()) {
4702 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
4706 AccessAnalysis::DepCandidates DependentAccesses;
4707 AccessAnalysis Accesses(DL, AA, DependentAccesses);
4709 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
4710 // multiple times on the same object. If the ptr is accessed twice, once
4711 // for read and once for write, it will only appear once (on the write
4712 // list). This is okay, since we are going to check for conflicts between
4713 // writes and between reads and writes, but not between reads and reads.
4716 ValueVector::iterator I, IE;
4717 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
4718 StoreInst *ST = cast<StoreInst>(*I);
4719 Value* Ptr = ST->getPointerOperand();
4721 if (isUniform(Ptr)) {
4724 << "write to a loop invariant address could not be vectorized");
4725 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4729 // If we did *not* see this pointer before, insert it to the read-write
4730 // list. At this phase it is only a 'write' list.
4731 if (Seen.insert(Ptr)) {
4734 AliasAnalysis::Location Loc = AA->getLocation(ST);
4735 // The TBAA metadata could have a control dependency on the predication
4736 // condition, so we cannot rely on it when determining whether or not we
4737 // need runtime pointer checks.
4738 if (blockNeedsPredication(ST->getParent()))
4739 Loc.AATags.TBAA = nullptr;
4741 Accesses.addStore(Loc);
4745 if (IsAnnotatedParallel) {
4747 << "LV: A loop annotated parallel, ignore memory dependency "
4752 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
4753 LoadInst *LD = cast<LoadInst>(*I);
4754 Value* Ptr = LD->getPointerOperand();
4755 // If we did *not* see this pointer before, insert it to the
4756 // read list. If we *did* see it before, then it is already in
4757 // the read-write list. This allows us to vectorize expressions
4758 // such as A[i] += x; Because the address of A[i] is a read-write
4759 // pointer. This only works if the index of A[i] is consecutive.
4760 // If the address of i is unknown (for example A[B[i]]) then we may
4761 // read a few words, modify, and write a few words, and some of the
4762 // words may be written to the same address.
4763 bool IsReadOnlyPtr = false;
4764 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
4766 IsReadOnlyPtr = true;
4769 AliasAnalysis::Location Loc = AA->getLocation(LD);
4770 // The TBAA metadata could have a control dependency on the predication
4771 // condition, so we cannot rely on it when determining whether or not we
4772 // need runtime pointer checks.
4773 if (blockNeedsPredication(LD->getParent()))
4774 Loc.AATags.TBAA = nullptr;
4776 Accesses.addLoad(Loc, IsReadOnlyPtr);
4779 // If we write (or read-write) to a single destination and there are no
4780 // other reads in this loop then is it safe to vectorize.
4781 if (NumReadWrites == 1 && NumReads == 0) {
4782 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
4786 // Build dependence sets and check whether we need a runtime pointer bounds
4788 Accesses.buildDependenceSets();
4789 bool NeedRTCheck = Accesses.isRTCheckNeeded();
4791 // Find pointers with computable bounds. We are going to use this information
4792 // to place a runtime bound check.
4793 unsigned NumComparisons = 0;
4794 bool CanDoRT = false;
4796 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
4799 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
4800 " pointer comparisons.\n");
4802 // If we only have one set of dependences to check pointers among we don't
4803 // need a runtime check.
4804 if (NumComparisons == 0 && NeedRTCheck)
4805 NeedRTCheck = false;
4807 // Check that we did not collect too many pointers or found an unsizeable
4809 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4815 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
4818 if (NeedRTCheck && !CanDoRT) {
4819 emitAnalysis(Report() << "cannot identify array bounds");
4820 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
4821 "the array bounds.\n");
4826 PtrRtCheck.Need = NeedRTCheck;
4828 bool CanVecMem = true;
4829 if (Accesses.isDependencyCheckNeeded()) {
4830 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
4831 CanVecMem = DepChecker.areDepsSafe(
4832 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
4833 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
4835 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
4836 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
4839 // Clear the dependency checks. We assume they are not needed.
4840 Accesses.resetDepChecks();
4843 PtrRtCheck.Need = true;
4845 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
4846 TheLoop, Strides, true);
4847 // Check that we did not collect too many pointers or found an unsizeable
4849 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4850 if (!CanDoRT && NumComparisons > 0)
4851 emitAnalysis(Report()
4852 << "cannot check memory dependencies at runtime");
4854 emitAnalysis(Report()
4855 << NumComparisons << " exceeds limit of "
4856 << RuntimeMemoryCheckThreshold
4857 << " dependent memory operations checked at runtime");
4858 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
4868 emitAnalysis(Report() << "unsafe dependent memory operations in loop");
4870 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
4871 " need a runtime memory check.\n");
4876 static bool hasMultipleUsesOf(Instruction *I,
4877 SmallPtrSet<Instruction *, 8> &Insts) {
4878 unsigned NumUses = 0;
4879 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
4880 if (Insts.count(dyn_cast<Instruction>(*Use)))
4889 static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
4890 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
4891 if (!Set.count(dyn_cast<Instruction>(*Use)))
4896 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
4897 ReductionKind Kind) {
4898 if (Phi->getNumIncomingValues() != 2)
4901 // Reduction variables are only found in the loop header block.
4902 if (Phi->getParent() != TheLoop->getHeader())
4905 // Obtain the reduction start value from the value that comes from the loop
4907 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
4909 // ExitInstruction is the single value which is used outside the loop.
4910 // We only allow for a single reduction value to be used outside the loop.
4911 // This includes users of the reduction, variables (which form a cycle
4912 // which ends in the phi node).
4913 Instruction *ExitInstruction = nullptr;
4914 // Indicates that we found a reduction operation in our scan.
4915 bool FoundReduxOp = false;
4917 // We start with the PHI node and scan for all of the users of this
4918 // instruction. All users must be instructions that can be used as reduction
4919 // variables (such as ADD). We must have a single out-of-block user. The cycle
4920 // must include the original PHI.
4921 bool FoundStartPHI = false;
4923 // To recognize min/max patterns formed by a icmp select sequence, we store
4924 // the number of instruction we saw from the recognized min/max pattern,
4925 // to make sure we only see exactly the two instructions.
4926 unsigned NumCmpSelectPatternInst = 0;
4927 ReductionInstDesc ReduxDesc(false, nullptr);
4929 SmallPtrSet<Instruction *, 8> VisitedInsts;
4930 SmallVector<Instruction *, 8> Worklist;
4931 Worklist.push_back(Phi);
4932 VisitedInsts.insert(Phi);
4934 // A value in the reduction can be used:
4935 // - By the reduction:
4936 // - Reduction operation:
4937 // - One use of reduction value (safe).
4938 // - Multiple use of reduction value (not safe).
4940 // - All uses of the PHI must be the reduction (safe).
4941 // - Otherwise, not safe.
4942 // - By one instruction outside of the loop (safe).
4943 // - By further instructions outside of the loop (not safe).
4944 // - By an instruction that is not part of the reduction (not safe).
4946 // * An instruction type other than PHI or the reduction operation.
4947 // * A PHI in the header other than the initial PHI.
4948 while (!Worklist.empty()) {
4949 Instruction *Cur = Worklist.back();
4950 Worklist.pop_back();
4953 // If the instruction has no users then this is a broken chain and can't be
4954 // a reduction variable.
4955 if (Cur->use_empty())
4958 bool IsAPhi = isa<PHINode>(Cur);
4960 // A header PHI use other than the original PHI.
4961 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
4964 // Reductions of instructions such as Div, and Sub is only possible if the
4965 // LHS is the reduction variable.
4966 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
4967 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
4968 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
4971 // Any reduction instruction must be of one of the allowed kinds.
4972 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
4973 if (!ReduxDesc.IsReduction)
4976 // A reduction operation must only have one use of the reduction value.
4977 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
4978 hasMultipleUsesOf(Cur, VisitedInsts))
4981 // All inputs to a PHI node must be a reduction value.
4982 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
4985 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
4986 isa<SelectInst>(Cur)))
4987 ++NumCmpSelectPatternInst;
4988 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
4989 isa<SelectInst>(Cur)))
4990 ++NumCmpSelectPatternInst;
4992 // Check whether we found a reduction operator.
4993 FoundReduxOp |= !IsAPhi;
4995 // Process users of current instruction. Push non-PHI nodes after PHI nodes
4996 // onto the stack. This way we are going to have seen all inputs to PHI
4997 // nodes once we get to them.
4998 SmallVector<Instruction *, 8> NonPHIs;
4999 SmallVector<Instruction *, 8> PHIs;
5000 for (User *U : Cur->users()) {
5001 Instruction *UI = cast<Instruction>(U);
5003 // Check if we found the exit user.
5004 BasicBlock *Parent = UI->getParent();
5005 if (!TheLoop->contains(Parent)) {
5006 // Exit if you find multiple outside users or if the header phi node is
5007 // being used. In this case the user uses the value of the previous
5008 // iteration, in which case we would loose "VF-1" iterations of the
5009 // reduction operation if we vectorize.
5010 if (ExitInstruction != nullptr || Cur == Phi)
5013 // The instruction used by an outside user must be the last instruction
5014 // before we feed back to the reduction phi. Otherwise, we loose VF-1
5015 // operations on the value.
5016 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
5019 ExitInstruction = Cur;
5023 // Process instructions only once (termination). Each reduction cycle
5024 // value must only be used once, except by phi nodes and min/max
5025 // reductions which are represented as a cmp followed by a select.
5026 ReductionInstDesc IgnoredVal(false, nullptr);
5027 if (VisitedInsts.insert(UI)) {
5028 if (isa<PHINode>(UI))
5031 NonPHIs.push_back(UI);
5032 } else if (!isa<PHINode>(UI) &&
5033 ((!isa<FCmpInst>(UI) &&
5034 !isa<ICmpInst>(UI) &&
5035 !isa<SelectInst>(UI)) ||
5036 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
5039 // Remember that we completed the cycle.
5041 FoundStartPHI = true;
5043 Worklist.append(PHIs.begin(), PHIs.end());
5044 Worklist.append(NonPHIs.begin(), NonPHIs.end());
5047 // This means we have seen one but not the other instruction of the
5048 // pattern or more than just a select and cmp.
5049 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
5050 NumCmpSelectPatternInst != 2)
5053 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
5056 // We found a reduction var if we have reached the original phi node and we
5057 // only have a single instruction with out-of-loop users.
5059 // This instruction is allowed to have out-of-loop users.
5060 AllowedExit.insert(ExitInstruction);
5062 // Save the description of this reduction variable.
5063 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
5064 ReduxDesc.MinMaxKind);
5065 Reductions[Phi] = RD;
5066 // We've ended the cycle. This is a reduction variable if we have an
5067 // outside user and it has a binary op.
5072 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
5073 /// pattern corresponding to a min(X, Y) or max(X, Y).
5074 LoopVectorizationLegality::ReductionInstDesc
5075 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
5076 ReductionInstDesc &Prev) {
5078 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
5079 "Expect a select instruction");
5080 Instruction *Cmp = nullptr;
5081 SelectInst *Select = nullptr;
5083 // We must handle the select(cmp()) as a single instruction. Advance to the
5085 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
5086 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
5087 return ReductionInstDesc(false, I);
5088 return ReductionInstDesc(Select, Prev.MinMaxKind);
5091 // Only handle single use cases for now.
5092 if (!(Select = dyn_cast<SelectInst>(I)))
5093 return ReductionInstDesc(false, I);
5094 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
5095 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
5096 return ReductionInstDesc(false, I);
5097 if (!Cmp->hasOneUse())
5098 return ReductionInstDesc(false, I);
5103 // Look for a min/max pattern.
5104 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5105 return ReductionInstDesc(Select, MRK_UIntMin);
5106 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5107 return ReductionInstDesc(Select, MRK_UIntMax);
5108 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5109 return ReductionInstDesc(Select, MRK_SIntMax);
5110 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5111 return ReductionInstDesc(Select, MRK_SIntMin);
5112 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5113 return ReductionInstDesc(Select, MRK_FloatMin);
5114 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5115 return ReductionInstDesc(Select, MRK_FloatMax);
5116 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5117 return ReductionInstDesc(Select, MRK_FloatMin);
5118 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5119 return ReductionInstDesc(Select, MRK_FloatMax);
5121 return ReductionInstDesc(false, I);
5124 LoopVectorizationLegality::ReductionInstDesc
5125 LoopVectorizationLegality::isReductionInstr(Instruction *I,
5127 ReductionInstDesc &Prev) {
5128 bool FP = I->getType()->isFloatingPointTy();
5129 bool FastMath = (FP && I->isCommutative() && I->isAssociative());
5130 switch (I->getOpcode()) {
5132 return ReductionInstDesc(false, I);
5133 case Instruction::PHI:
5134 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
5135 Kind != RK_FloatMinMax))
5136 return ReductionInstDesc(false, I);
5137 return ReductionInstDesc(I, Prev.MinMaxKind);
5138 case Instruction::Sub:
5139 case Instruction::Add:
5140 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
5141 case Instruction::Mul:
5142 return ReductionInstDesc(Kind == RK_IntegerMult, I);
5143 case Instruction::And:
5144 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
5145 case Instruction::Or:
5146 return ReductionInstDesc(Kind == RK_IntegerOr, I);
5147 case Instruction::Xor:
5148 return ReductionInstDesc(Kind == RK_IntegerXor, I);
5149 case Instruction::FMul:
5150 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
5151 case Instruction::FAdd:
5152 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
5153 case Instruction::FCmp:
5154 case Instruction::ICmp:
5155 case Instruction::Select:
5156 if (Kind != RK_IntegerMinMax &&
5157 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
5158 return ReductionInstDesc(false, I);
5159 return isMinMaxSelectCmpPattern(I, Prev);
5163 LoopVectorizationLegality::InductionKind
5164 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
5165 Type *PhiTy = Phi->getType();
5166 // We only handle integer and pointer inductions variables.
5167 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
5168 return IK_NoInduction;
5170 // Check that the PHI is consecutive.
5171 const SCEV *PhiScev = SE->getSCEV(Phi);
5172 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
5174 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
5175 return IK_NoInduction;
5177 const SCEV *Step = AR->getStepRecurrence(*SE);
5179 // Integer inductions need to have a stride of one.
5180 if (PhiTy->isIntegerTy()) {
5182 return IK_IntInduction;
5183 if (Step->isAllOnesValue())
5184 return IK_ReverseIntInduction;
5185 return IK_NoInduction;
5188 // Calculate the pointer stride and check if it is consecutive.
5189 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5191 return IK_NoInduction;
5193 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
5194 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
5195 if (C->getValue()->equalsInt(Size))
5196 return IK_PtrInduction;
5197 else if (C->getValue()->equalsInt(0 - Size))
5198 return IK_ReversePtrInduction;
5200 return IK_NoInduction;
5203 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
5204 Value *In0 = const_cast<Value*>(V);
5205 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
5209 return Inductions.count(PN);
5212 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
5213 assert(TheLoop->contains(BB) && "Unknown block used");
5215 // Blocks that do not dominate the latch need predication.
5216 BasicBlock* Latch = TheLoop->getLoopLatch();
5217 return !DT->dominates(BB, Latch);
5220 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
5221 SmallPtrSet<Value *, 8>& SafePtrs) {
5222 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5223 // We might be able to hoist the load.
5224 if (it->mayReadFromMemory()) {
5225 LoadInst *LI = dyn_cast<LoadInst>(it);
5226 if (!LI || !SafePtrs.count(LI->getPointerOperand()))
5230 // We don't predicate stores at the moment.
5231 if (it->mayWriteToMemory()) {
5232 StoreInst *SI = dyn_cast<StoreInst>(it);
5233 // We only support predication of stores in basic blocks with one
5235 if (!SI || ++NumPredStores > NumberOfStoresToPredicate ||
5236 !SafePtrs.count(SI->getPointerOperand()) ||
5237 !SI->getParent()->getSinglePredecessor())
5243 // Check that we don't have a constant expression that can trap as operand.
5244 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
5246 if (Constant *C = dyn_cast<Constant>(*OI))
5251 // The instructions below can trap.
5252 switch (it->getOpcode()) {
5254 case Instruction::UDiv:
5255 case Instruction::SDiv:
5256 case Instruction::URem:
5257 case Instruction::SRem:
5265 LoopVectorizationCostModel::VectorizationFactor
5266 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
5268 bool ForceVectorization) {
5269 // Width 1 means no vectorize
5270 VectorizationFactor Factor = { 1U, 0U };
5271 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
5272 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
5276 if (!EnableCondStoresVectorization && Legal->NumPredStores) {
5277 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
5281 // Find the trip count.
5282 unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
5283 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
5285 unsigned WidestType = getWidestType();
5286 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
5287 unsigned MaxSafeDepDist = -1U;
5288 if (Legal->getMaxSafeDepDistBytes() != -1U)
5289 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
5290 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
5291 WidestRegister : MaxSafeDepDist);
5292 unsigned MaxVectorSize = WidestRegister / WidestType;
5293 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
5294 DEBUG(dbgs() << "LV: The Widest register is: "
5295 << WidestRegister << " bits.\n");
5297 if (MaxVectorSize == 0) {
5298 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
5302 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
5303 " into one vector!");
5305 unsigned VF = MaxVectorSize;
5307 // If we optimize the program for size, avoid creating the tail loop.
5309 // If we are unable to calculate the trip count then don't try to vectorize.
5311 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5315 // Find the maximum SIMD width that can fit within the trip count.
5316 VF = TC % MaxVectorSize;
5321 // If the trip count that we found modulo the vectorization factor is not
5322 // zero then we require a tail.
5324 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5330 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
5331 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
5333 Factor.Width = UserVF;
5337 float Cost = expectedCost(1);
5339 const float ScalarCost = Cost;
5342 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
5344 // Ignore scalar width, because the user explicitly wants vectorization.
5345 if (ForceVectorization && VF > 1) {
5347 Cost = expectedCost(Width) / (float)Width;
5350 for (unsigned i=2; i <= VF; i*=2) {
5351 // Notice that the vector loop needs to be executed less times, so
5352 // we need to divide the cost of the vector loops by the width of
5353 // the vector elements.
5354 float VectorCost = expectedCost(i) / (float)i;
5355 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
5356 (int)VectorCost << ".\n");
5357 if (VectorCost < Cost) {
5363 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
5364 << "LV: Vectorization seems to be not beneficial, "
5365 << "but was forced by a user.\n");
5366 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
5367 Factor.Width = Width;
5368 Factor.Cost = Width * Cost;
5372 unsigned LoopVectorizationCostModel::getWidestType() {
5373 unsigned MaxWidth = 8;
5376 for (Loop::block_iterator bb = TheLoop->block_begin(),
5377 be = TheLoop->block_end(); bb != be; ++bb) {
5378 BasicBlock *BB = *bb;
5380 // For each instruction in the loop.
5381 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5382 Type *T = it->getType();
5384 // Only examine Loads, Stores and PHINodes.
5385 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
5388 // Examine PHI nodes that are reduction variables.
5389 if (PHINode *PN = dyn_cast<PHINode>(it))
5390 if (!Legal->getReductionVars()->count(PN))
5393 // Examine the stored values.
5394 if (StoreInst *ST = dyn_cast<StoreInst>(it))
5395 T = ST->getValueOperand()->getType();
5397 // Ignore loaded pointer types and stored pointer types that are not
5398 // consecutive. However, we do want to take consecutive stores/loads of
5399 // pointer vectors into account.
5400 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
5403 MaxWidth = std::max(MaxWidth,
5404 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
5412 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
5415 unsigned LoopCost) {
5417 // -- The unroll heuristics --
5418 // We unroll the loop in order to expose ILP and reduce the loop overhead.
5419 // There are many micro-architectural considerations that we can't predict
5420 // at this level. For example frontend pressure (on decode or fetch) due to
5421 // code size, or the number and capabilities of the execution ports.
5423 // We use the following heuristics to select the unroll factor:
5424 // 1. If the code has reductions the we unroll in order to break the cross
5425 // iteration dependency.
5426 // 2. If the loop is really small then we unroll in order to reduce the loop
5428 // 3. We don't unroll if we think that we will spill registers to memory due
5429 // to the increased register pressure.
5431 // Use the user preference, unless 'auto' is selected.
5435 // When we optimize for size we don't unroll.
5439 // We used the distance for the unroll factor.
5440 if (Legal->getMaxSafeDepDistBytes() != -1U)
5443 // Do not unroll loops with a relatively small trip count.
5444 unsigned TC = SE->getSmallConstantTripCount(TheLoop,
5445 TheLoop->getLoopLatch());
5446 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
5449 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
5450 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
5454 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
5455 TargetNumRegisters = ForceTargetNumScalarRegs;
5457 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
5458 TargetNumRegisters = ForceTargetNumVectorRegs;
5461 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
5462 // We divide by these constants so assume that we have at least one
5463 // instruction that uses at least one register.
5464 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
5465 R.NumInstructions = std::max(R.NumInstructions, 1U);
5467 // We calculate the unroll factor using the following formula.
5468 // Subtract the number of loop invariants from the number of available
5469 // registers. These registers are used by all of the unrolled instances.
5470 // Next, divide the remaining registers by the number of registers that is
5471 // required by the loop, in order to estimate how many parallel instances
5472 // fit without causing spills. All of this is rounded down if necessary to be
5473 // a power of two. We want power of two unroll factors to simplify any
5474 // addressing operations or alignment considerations.
5475 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
5478 // Don't count the induction variable as unrolled.
5479 if (EnableIndVarRegisterHeur)
5480 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
5481 std::max(1U, (R.MaxLocalUsers - 1)));
5483 // Clamp the unroll factor ranges to reasonable factors.
5484 unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
5486 // Check if the user has overridden the unroll max.
5488 if (ForceTargetMaxScalarUnrollFactor.getNumOccurrences() > 0)
5489 MaxUnrollSize = ForceTargetMaxScalarUnrollFactor;
5491 if (ForceTargetMaxVectorUnrollFactor.getNumOccurrences() > 0)
5492 MaxUnrollSize = ForceTargetMaxVectorUnrollFactor;
5495 // If we did not calculate the cost for VF (because the user selected the VF)
5496 // then we calculate the cost of VF here.
5498 LoopCost = expectedCost(VF);
5500 // Clamp the calculated UF to be between the 1 and the max unroll factor
5501 // that the target allows.
5502 if (UF > MaxUnrollSize)
5507 // Unroll if we vectorized this loop and there is a reduction that could
5508 // benefit from unrolling.
5509 if (VF > 1 && Legal->getReductionVars()->size()) {
5510 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
5514 // Note that if we've already vectorized the loop we will have done the
5515 // runtime check and so unrolling won't require further checks.
5516 bool UnrollingRequiresRuntimePointerCheck =
5517 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
5519 // We want to unroll small loops in order to reduce the loop overhead and
5520 // potentially expose ILP opportunities.
5521 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
5522 if (!UnrollingRequiresRuntimePointerCheck &&
5523 LoopCost < SmallLoopCost) {
5524 // We assume that the cost overhead is 1 and we use the cost model
5525 // to estimate the cost of the loop and unroll until the cost of the
5526 // loop overhead is about 5% of the cost of the loop.
5527 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
5529 // Unroll until store/load ports (estimated by max unroll factor) are
5531 unsigned StoresUF = UF / (Legal->NumStores ? Legal->NumStores : 1);
5532 unsigned LoadsUF = UF / (Legal->NumLoads ? Legal->NumLoads : 1);
5534 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
5535 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
5536 return std::max(StoresUF, LoadsUF);
5539 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
5543 DEBUG(dbgs() << "LV: Not Unrolling.\n");
5547 LoopVectorizationCostModel::RegisterUsage
5548 LoopVectorizationCostModel::calculateRegisterUsage() {
5549 // This function calculates the register usage by measuring the highest number
5550 // of values that are alive at a single location. Obviously, this is a very
5551 // rough estimation. We scan the loop in a topological order in order and
5552 // assign a number to each instruction. We use RPO to ensure that defs are
5553 // met before their users. We assume that each instruction that has in-loop
5554 // users starts an interval. We record every time that an in-loop value is
5555 // used, so we have a list of the first and last occurrences of each
5556 // instruction. Next, we transpose this data structure into a multi map that
5557 // holds the list of intervals that *end* at a specific location. This multi
5558 // map allows us to perform a linear search. We scan the instructions linearly
5559 // and record each time that a new interval starts, by placing it in a set.
5560 // If we find this value in the multi-map then we remove it from the set.
5561 // The max register usage is the maximum size of the set.
5562 // We also search for instructions that are defined outside the loop, but are
5563 // used inside the loop. We need this number separately from the max-interval
5564 // usage number because when we unroll, loop-invariant values do not take
5566 LoopBlocksDFS DFS(TheLoop);
5570 R.NumInstructions = 0;
5572 // Each 'key' in the map opens a new interval. The values
5573 // of the map are the index of the 'last seen' usage of the
5574 // instruction that is the key.
5575 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5576 // Maps instruction to its index.
5577 DenseMap<unsigned, Instruction*> IdxToInstr;
5578 // Marks the end of each interval.
5579 IntervalMap EndPoint;
5580 // Saves the list of instruction indices that are used in the loop.
5581 SmallSet<Instruction*, 8> Ends;
5582 // Saves the list of values that are used in the loop but are
5583 // defined outside the loop, such as arguments and constants.
5584 SmallPtrSet<Value*, 8> LoopInvariants;
5587 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5588 be = DFS.endRPO(); bb != be; ++bb) {
5589 R.NumInstructions += (*bb)->size();
5590 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
5592 Instruction *I = it;
5593 IdxToInstr[Index++] = I;
5595 // Save the end location of each USE.
5596 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
5597 Value *U = I->getOperand(i);
5598 Instruction *Instr = dyn_cast<Instruction>(U);
5600 // Ignore non-instruction values such as arguments, constants, etc.
5601 if (!Instr) continue;
5603 // If this instruction is outside the loop then record it and continue.
5604 if (!TheLoop->contains(Instr)) {
5605 LoopInvariants.insert(Instr);
5609 // Overwrite previous end points.
5610 EndPoint[Instr] = Index;
5616 // Saves the list of intervals that end with the index in 'key'.
5617 typedef SmallVector<Instruction*, 2> InstrList;
5618 DenseMap<unsigned, InstrList> TransposeEnds;
5620 // Transpose the EndPoints to a list of values that end at each index.
5621 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5623 TransposeEnds[it->second].push_back(it->first);
5625 SmallSet<Instruction*, 8> OpenIntervals;
5626 unsigned MaxUsage = 0;
5629 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5630 for (unsigned int i = 0; i < Index; ++i) {
5631 Instruction *I = IdxToInstr[i];
5632 // Ignore instructions that are never used within the loop.
5633 if (!Ends.count(I)) continue;
5635 // Remove all of the instructions that end at this location.
5636 InstrList &List = TransposeEnds[i];
5637 for (unsigned int j=0, e = List.size(); j < e; ++j)
5638 OpenIntervals.erase(List[j]);
5640 // Count the number of live interals.
5641 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5643 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5644 OpenIntervals.size() << '\n');
5646 // Add the current instruction to the list of open intervals.
5647 OpenIntervals.insert(I);
5650 unsigned Invariant = LoopInvariants.size();
5651 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5652 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5653 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5655 R.LoopInvariantRegs = Invariant;
5656 R.MaxLocalUsers = MaxUsage;
5660 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5664 for (Loop::block_iterator bb = TheLoop->block_begin(),
5665 be = TheLoop->block_end(); bb != be; ++bb) {
5666 unsigned BlockCost = 0;
5667 BasicBlock *BB = *bb;
5669 // For each instruction in the old loop.
5670 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5671 // Skip dbg intrinsics.
5672 if (isa<DbgInfoIntrinsic>(it))
5675 unsigned C = getInstructionCost(it, VF);
5677 // Check if we should override the cost.
5678 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5679 C = ForceTargetInstructionCost;
5682 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5683 VF << " For instruction: " << *it << '\n');
5686 // We assume that if-converted blocks have a 50% chance of being executed.
5687 // When the code is scalar then some of the blocks are avoided due to CF.
5688 // When the code is vectorized we execute all code paths.
5689 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5698 /// \brief Check whether the address computation for a non-consecutive memory
5699 /// access looks like an unlikely candidate for being merged into the indexing
5702 /// We look for a GEP which has one index that is an induction variable and all
5703 /// other indices are loop invariant. If the stride of this access is also
5704 /// within a small bound we decide that this address computation can likely be
5705 /// merged into the addressing mode.
5706 /// In all other cases, we identify the address computation as complex.
5707 static bool isLikelyComplexAddressComputation(Value *Ptr,
5708 LoopVectorizationLegality *Legal,
5709 ScalarEvolution *SE,
5710 const Loop *TheLoop) {
5711 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5715 // We are looking for a gep with all loop invariant indices except for one
5716 // which should be an induction variable.
5717 unsigned NumOperands = Gep->getNumOperands();
5718 for (unsigned i = 1; i < NumOperands; ++i) {
5719 Value *Opd = Gep->getOperand(i);
5720 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5721 !Legal->isInductionVariable(Opd))
5725 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5726 // can likely be merged into the address computation.
5727 unsigned MaxMergeDistance = 64;
5729 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5733 // Check the step is constant.
5734 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5735 // Calculate the pointer stride and check if it is consecutive.
5736 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5740 const APInt &APStepVal = C->getValue()->getValue();
5742 // Huge step value - give up.
5743 if (APStepVal.getBitWidth() > 64)
5746 int64_t StepVal = APStepVal.getSExtValue();
5748 return StepVal > MaxMergeDistance;
5751 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5752 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5758 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5759 // If we know that this instruction will remain uniform, check the cost of
5760 // the scalar version.
5761 if (Legal->isUniformAfterVectorization(I))
5764 Type *RetTy = I->getType();
5765 Type *VectorTy = ToVectorTy(RetTy, VF);
5767 // TODO: We need to estimate the cost of intrinsic calls.
5768 switch (I->getOpcode()) {
5769 case Instruction::GetElementPtr:
5770 // We mark this instruction as zero-cost because the cost of GEPs in
5771 // vectorized code depends on whether the corresponding memory instruction
5772 // is scalarized or not. Therefore, we handle GEPs with the memory
5773 // instruction cost.
5775 case Instruction::Br: {
5776 return TTI.getCFInstrCost(I->getOpcode());
5778 case Instruction::PHI:
5779 //TODO: IF-converted IFs become selects.
5781 case Instruction::Add:
5782 case Instruction::FAdd:
5783 case Instruction::Sub:
5784 case Instruction::FSub:
5785 case Instruction::Mul:
5786 case Instruction::FMul:
5787 case Instruction::UDiv:
5788 case Instruction::SDiv:
5789 case Instruction::FDiv:
5790 case Instruction::URem:
5791 case Instruction::SRem:
5792 case Instruction::FRem:
5793 case Instruction::Shl:
5794 case Instruction::LShr:
5795 case Instruction::AShr:
5796 case Instruction::And:
5797 case Instruction::Or:
5798 case Instruction::Xor: {
5799 // Since we will replace the stride by 1 the multiplication should go away.
5800 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5802 // Certain instructions can be cheaper to vectorize if they have a constant
5803 // second vector operand. One example of this are shifts on x86.
5804 TargetTransformInfo::OperandValueKind Op1VK =
5805 TargetTransformInfo::OK_AnyValue;
5806 TargetTransformInfo::OperandValueKind Op2VK =
5807 TargetTransformInfo::OK_AnyValue;
5808 Value *Op2 = I->getOperand(1);
5810 // Check for a splat of a constant or for a non uniform vector of constants.
5811 if (isa<ConstantInt>(Op2))
5812 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5813 else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5814 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5815 if (cast<Constant>(Op2)->getSplatValue() != nullptr)
5816 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5819 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
5821 case Instruction::Select: {
5822 SelectInst *SI = cast<SelectInst>(I);
5823 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5824 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5825 Type *CondTy = SI->getCondition()->getType();
5827 CondTy = VectorType::get(CondTy, VF);
5829 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5831 case Instruction::ICmp:
5832 case Instruction::FCmp: {
5833 Type *ValTy = I->getOperand(0)->getType();
5834 VectorTy = ToVectorTy(ValTy, VF);
5835 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5837 case Instruction::Store:
5838 case Instruction::Load: {
5839 StoreInst *SI = dyn_cast<StoreInst>(I);
5840 LoadInst *LI = dyn_cast<LoadInst>(I);
5841 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5843 VectorTy = ToVectorTy(ValTy, VF);
5845 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5846 unsigned AS = SI ? SI->getPointerAddressSpace() :
5847 LI->getPointerAddressSpace();
5848 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5849 // We add the cost of address computation here instead of with the gep
5850 // instruction because only here we know whether the operation is
5853 return TTI.getAddressComputationCost(VectorTy) +
5854 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5856 // Scalarized loads/stores.
5857 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5858 bool Reverse = ConsecutiveStride < 0;
5859 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
5860 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
5861 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5862 bool IsComplexComputation =
5863 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5865 // The cost of extracting from the value vector and pointer vector.
5866 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5867 for (unsigned i = 0; i < VF; ++i) {
5868 // The cost of extracting the pointer operand.
5869 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5870 // In case of STORE, the cost of ExtractElement from the vector.
5871 // In case of LOAD, the cost of InsertElement into the returned
5873 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5874 Instruction::InsertElement,
5878 // The cost of the scalar loads/stores.
5879 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5880 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5885 // Wide load/stores.
5886 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5887 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5890 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5894 case Instruction::ZExt:
5895 case Instruction::SExt:
5896 case Instruction::FPToUI:
5897 case Instruction::FPToSI:
5898 case Instruction::FPExt:
5899 case Instruction::PtrToInt:
5900 case Instruction::IntToPtr:
5901 case Instruction::SIToFP:
5902 case Instruction::UIToFP:
5903 case Instruction::Trunc:
5904 case Instruction::FPTrunc:
5905 case Instruction::BitCast: {
5906 // We optimize the truncation of induction variable.
5907 // The cost of these is the same as the scalar operation.
5908 if (I->getOpcode() == Instruction::Trunc &&
5909 Legal->isInductionVariable(I->getOperand(0)))
5910 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
5911 I->getOperand(0)->getType());
5913 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
5914 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
5916 case Instruction::Call: {
5917 CallInst *CI = cast<CallInst>(I);
5918 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
5919 assert(ID && "Not an intrinsic call!");
5920 Type *RetTy = ToVectorTy(CI->getType(), VF);
5921 SmallVector<Type*, 4> Tys;
5922 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
5923 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
5924 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
5927 // We are scalarizing the instruction. Return the cost of the scalar
5928 // instruction, plus the cost of insert and extract into vector
5929 // elements, times the vector width.
5932 if (!RetTy->isVoidTy() && VF != 1) {
5933 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
5935 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
5938 // The cost of inserting the results plus extracting each one of the
5940 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
5943 // The cost of executing VF copies of the scalar instruction. This opcode
5944 // is unknown. Assume that it is the same as 'mul'.
5945 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
5951 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
5952 if (Scalar->isVoidTy() || VF == 1)
5954 return VectorType::get(Scalar, VF);
5957 char LoopVectorize::ID = 0;
5958 static const char lv_name[] = "Loop Vectorization";
5959 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5960 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
5961 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
5962 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
5963 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5964 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5965 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5966 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
5967 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5968 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5971 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5972 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5976 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5977 // Check for a store.
5978 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5979 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5981 // Check for a load.
5982 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5983 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5989 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5990 bool IfPredicateStore) {
5991 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5992 // Holds vector parameters or scalars, in case of uniform vals.
5993 SmallVector<VectorParts, 4> Params;
5995 setDebugLocFromInst(Builder, Instr);
5997 // Find all of the vectorized parameters.
5998 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5999 Value *SrcOp = Instr->getOperand(op);
6001 // If we are accessing the old induction variable, use the new one.
6002 if (SrcOp == OldInduction) {
6003 Params.push_back(getVectorValue(SrcOp));
6007 // Try using previously calculated values.
6008 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
6010 // If the src is an instruction that appeared earlier in the basic block
6011 // then it should already be vectorized.
6012 if (SrcInst && OrigLoop->contains(SrcInst)) {
6013 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
6014 // The parameter is a vector value from earlier.
6015 Params.push_back(WidenMap.get(SrcInst));
6017 // The parameter is a scalar from outside the loop. Maybe even a constant.
6018 VectorParts Scalars;
6019 Scalars.append(UF, SrcOp);
6020 Params.push_back(Scalars);
6024 assert(Params.size() == Instr->getNumOperands() &&
6025 "Invalid number of operands");
6027 // Does this instruction return a value ?
6028 bool IsVoidRetTy = Instr->getType()->isVoidTy();
6030 Value *UndefVec = IsVoidRetTy ? nullptr :
6031 UndefValue::get(Instr->getType());
6032 // Create a new entry in the WidenMap and initialize it to Undef or Null.
6033 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
6035 Instruction *InsertPt = Builder.GetInsertPoint();
6036 BasicBlock *IfBlock = Builder.GetInsertBlock();
6037 BasicBlock *CondBlock = nullptr;
6040 Loop *VectorLp = nullptr;
6041 if (IfPredicateStore) {
6042 assert(Instr->getParent()->getSinglePredecessor() &&
6043 "Only support single predecessor blocks");
6044 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
6045 Instr->getParent());
6046 VectorLp = LI->getLoopFor(IfBlock);
6047 assert(VectorLp && "Must have a loop for this block");
6050 // For each vector unroll 'part':
6051 for (unsigned Part = 0; Part < UF; ++Part) {
6052 // For each scalar that we create:
6054 // Start an "if (pred) a[i] = ..." block.
6055 Value *Cmp = nullptr;
6056 if (IfPredicateStore) {
6057 if (Cond[Part]->getType()->isVectorTy())
6059 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
6060 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
6061 ConstantInt::get(Cond[Part]->getType(), 1));
6062 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
6063 LoopVectorBody.push_back(CondBlock);
6064 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
6065 // Update Builder with newly created basic block.
6066 Builder.SetInsertPoint(InsertPt);
6069 Instruction *Cloned = Instr->clone();
6071 Cloned->setName(Instr->getName() + ".cloned");
6072 // Replace the operands of the cloned instructions with extracted scalars.
6073 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6074 Value *Op = Params[op][Part];
6075 Cloned->setOperand(op, Op);
6078 // Place the cloned scalar in the new loop.
6079 Builder.Insert(Cloned);
6081 // If the original scalar returns a value we need to place it in a vector
6082 // so that future users will be able to use it.
6084 VecResults[Part] = Cloned;
6087 if (IfPredicateStore) {
6088 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
6089 LoopVectorBody.push_back(NewIfBlock);
6090 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
6091 Builder.SetInsertPoint(InsertPt);
6092 Instruction *OldBr = IfBlock->getTerminator();
6093 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
6094 OldBr->eraseFromParent();
6095 IfBlock = NewIfBlock;
6100 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
6101 StoreInst *SI = dyn_cast<StoreInst>(Instr);
6102 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
6104 return scalarizeInstruction(Instr, IfPredicateStore);
6107 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
6111 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
6115 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
6117 // When unrolling and the VF is 1, we only need to add a simple scalar.
6118 Type *ITy = Val->getType();
6119 assert(!ITy->isVectorTy() && "Val must be a scalar");
6120 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
6121 return Builder.CreateAdd(Val, C, "induction");